System and method for discriminating and gradually controlling automated-actor traffic based on comparison between network behavior contracts and observed behavior
The system uses flow metadata to identify and manage unauthorized network communications, reducing false positives and business risks by applying controlled responses with minimal operational impact.
Patent Information
- Application Number
- KR1020260189430
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2026-10-02
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2046-10-02
AI Technical Summary
Existing methods struggle to distinguish between authorized and unauthorized network communications by automation entities, such as AI agents, without decoding the communication content, leading to false positives, contamination of approval criteria, and risks of business interruption due to overzealous control responses.
A system that identifies automation subjects using flow metadata, manages contract deviations as candidate states, and applies temporary control overlays without overwriting standard policies, allowing for stepwise control and verification of adaptive behaviors.
Reduces the burden of protecting personal information, minimizes false positives, and reduces the risk of business interruption by applying controlled responses with minimal impact on operations, enabling accurate identification and management of unauthorized network communications.
Smart Images

Figure 112026122161948-PAT00001_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to network security, control of artificial intelligence agents, and network digital twins.
[0002] More specifically, the present invention relates to a system and method for identifying an automation entity that generates network communication without individual human operation, such as an artificial intelligence agent, bot, script, browser extension, scheduled task, business automation tool, and external automated reconnaissance / infiltration tool, using flow metadata that has not decoded the communication content, and managing contract deviations as candidate states separated from the approval criteria state by comparing the approved network behavior contract and the observed behavior for each automation entity.
[0003] Furthermore, the present invention relates to a technology for applying control stages arranged in order of business impact as a temporary control overlay that does not overwrite a stored standard policy, and for updating control stages and network behavior contracts based on the results of verifying and distinguishing the requested control state, the actually applied control state, the control effect, and the adaptive behavior of the automation agent. Background Technology
[0004] Artificial intelligence agents utilizing large-scale language models access external AI services, remote tool servers, internal file servers, and business systems on behalf of users. These agents communicate faster and more continuously than humans, transmit unintended data externally, or navigate the internal network by modifying their behavior based on instructions contained in documents, web pages, or tool responses.
[0005] External attackers can also utilize artificial intelligence to perform automated reconnaissance and infiltration by rapidly changing paths and inputs based on responses. Therefore, it is necessary for organizational networks to distinguish between human communications and communications by automated entities, and to verify whether these entities are operating within authorized scopes.
[0006] Conventionally, methods were used to prevent the external transmission of sensitive information by inspecting transmission content, such as prompts or uploaded files. However, since this method is based on the premise of viewing communication content, it places a heavy burden on protecting personal information and communication secrets, and is difficult to apply to end-to-end encryption, self-hosting models, and calls made via application programming interfaces.
[0007] Methods for visualizing unauthorized AI usage by using a list of destinations or application signatures of known generative AI services in network traffic are known, and structures for detecting AI usage as behavioral fingerprints by analyzing network traffic, application programming interface calls, browser extensions, etc. have also been proposed (Non-patent Literature 3). However, these methods mainly remain at the level of visualization and listing, and do not address stepwise control after detection and verification of the results.
[0008] There are also known methods in which a sending proxy is placed in the agent execution environment to allow only allowed domains to pass through and require approval for unknown destinations, and in which the network restricts the device's communication according to a list of allowed communications declared by the device manufacturer (Non-patent Literature 1). However, since these methods are based on the premise that the execution environment or the device is registered in the policy path, they fail to identify automation entities that arise outside the policy path, such as agents or browser extensions arbitrarily installed by the user, and remain at a single allow or block judgment regarding violations of the allow list.
[0009] Techniques for detecting mechanical communication forms, such as periodic beacons, as flow-level features, and techniques for classifying clients accessing web services based on behavioral features and learning models to perform blocking, rate limiting, or verification requests against bots (Patent Document 1, Patent Document 2) are also known. Recently, it has been reported that the communication of an agent performing a remote tool call can appear similar to a command and control beacon at the flow level (Non-patent Document 2).
[0010] However, these techniques generally remain at the level of classification based on individual requests or single services, and do not distinguish between organizationally approved and unapproved business automation through contracts based on the automation entity. As a result, approved backups, synchronizations, or business automations may be repeatedly classified as abnormal, or conversely, the behavior of unapproved automation may be trained on normal standards, thereby contaminating the criteria.
[0011] Furthermore, if the response after detection is primarily block-oriented, automated responses are prone to operating with their disabled due to the risk of suspending even authorized operations. Additionally, if a response is deemed successful solely based on the fact that a response command has been received, it misses bypasses such as the automation entity changing the retry interval after control, switching to an alternative destination, or distributing to different execution units.
[0012] Although a technique (Patent Document 3) is known in which network policies are created and distributed to each enforcement branch, and the enforcement branch verifies the suitability of the policies and applies them, a configuration is not presented in which contract deviations by automation entity are managed as candidate states separated from approval criteria, controls are applied as temporary overlays in order of business impact, and controls and contracts are updated according to the verification results.
[0013] In Korea, there are known techniques for calculating anomalies and automatically updating security policy tables by learning transmission and reception cycles, traffic size, destinations, and session durations by terminal and time period (Patent Document 4); techniques for restricting and blocking requests by using a learning model to determine whether a request is a macro or bot based on the time interval between behavioral events, request frequency, and request order, and for retraining the model by analyzing reconnection attempts and behavioral changes after blocking (Patent Document 5); and techniques for determining malicious bots by sequentially applying a whitelist, reference behavior, attack signatures, and TLS fingerprints (Patent Document 6).
[0014] In addition, there are known techniques for dropping packets if there is no authorization flow defined by application identification information and destination address / port (Patent Document 7), techniques for applying a temporary security policy only for a limited time or number of times and automatically returning to the previous policy after the period has elapsed (Patent Document 8), and techniques for evaluating and replacing the policy based on the deviation between prediction and actual measurement by feeding back the operational results after the policy is applied (Patent Document 9).
[0015] However, these techniques do not present a configuration that places the automation principal as a node in the approval standard twin and manages it as a principal-specific contract version, a configuration that separates contract exits into candidate branches bound to the standard version, or a configuration that distinguishes changes in the retry interval, alternative destination switching, or dispersion to other attribution units exhibited by the automation principal after control from the control effect and reflects this in raising the control level or expanding the scope of application. Prior art literature
[0016] U.S. Patent No. 9,942,214 (Automated agent detection utilizing non-CAPTCHA methods) U.S. Patent No. 12,430,254 (Web bot detection via behavioral analysis and machine learning) U.S. Patent No. 11,252,038 (Network agent for generating platform-specific network policies) Korean Registered Patent Publication No. 10-3009930 (AI-based generation network security policy learning and anomaly detection system) Korean Registered Patent Publication No. 10-3008759 (Method and system for blocking macrobot ticketing using AI-based user behavior patterns) Korean Registered Patent Publication No. 10-2909534 (Abnormal crawling determination system) Korean Registered Patent Publication No. 10-2309116 (System for controlling network access of a data flow-based application and method related thereto) Korean Published Patent Publication No. 10-2003-0003593 (Specific during a limitation Network security device and network security method capable of applying security policies) Korean Registered Patent Publication No. 10-3011029 (Closed-loop based intelligent network predictive control system using policy reliability evaluation and decision-making model)
[0017] E. Lear, R. Droms, D. Romascanu, "Manufacturer Usage Description Specification", IETF RFC 8520, 2019."When Agents Look Like Beacons: NIDS Evasion by Model Context Protocol Traffic", arXiv:2609.19091, 2026."Governing at Machine Speed: An Adaptive Intelligence Architecture for Real-Time AI Policy Enforcement", arXiv:2609.13466, 2026. The problem to be solved
[0018] The present invention aims to solve the above-mentioned problems by providing a system and method for identifying an automation subject as an independent observation target of a network digital twin using only flow metadata without decoding communication content.
[0019] Furthermore, the present invention aims to distinguish between automations approved by network behavior contracts for each automation subject and those that are not approved or have withdrawn, and to accumulate evidence by separating contract withdrawals into a candidate state rather than directly reflecting them in the approval standard state.
[0020] In addition, the present invention aims to ensure the safety of automatic response by applying a temporary overlay that does not overwrite the stored standard policy, starting from a control with a small business impact, and by preventing the safety limit from raising the control level.
[0021] Furthermore, the present invention aims to distinguish the requested state, actual application state, and effect of control, and to separately verify the adaptive behavior of the automation agent to detect false successes and circumventions, and thereby renew the control stage and contract as a result.
[0022] The problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned problems can be clearly understood by a person skilled in the art from the description below. means of solving the problem
[0023] An automated subject network behavior control system according to one aspect of the present invention for solving the above problem includes a memory for storing program commands and one or more processors for executing said program commands.
[0024] The above one or more processors receive flow metadata that has not been decoded from one or more network observers and can group the flow metadata into attribution units including at least one of a terminal identifier, an attribution entity identifier, a pseudonym user identifier, and a client fingerprint.
[0025] The above one or more processors can calculate behavioral characteristics for each attribution unit, including at least two of call interval distribution, clustering, fan-out, transmission and reception ratio, and response delay, and determine the possibility of automation indicating the possibility that the communication of the attribution unit was generated without individual human operation based on the behavioral characteristics, thereby identifying one or more automation agents.
[0026] The above one or more processors can manage an approval standard network digital twin including the automation principal as a node, and a network behavior contract including at least one of an allowed destination, allowed data class, rate envelope, and allowed time zone, and a contract version for each automation principal.
[0027] The above one or more processors may compare observed behavioral features, destinations, and data movements with the network behavior contract to determine at least one contract deviation among unregistered principal, destination deviation, data class deviation, velocity deviation, and time zone deviation, and generate the contract deviation as an automated principal candidate branch including a reference version identifier and a change delta without directly merging it into the approval reference network digital twin.
[0028] Based on the type of contract deviation and the possibility of automation, the above one or more processors may select one of a plurality of control steps arranged in order of business impact, including at least three of maintaining observation, deceleration within a speed limit, concurrency or session limit, destination category limit, confirmation request, and blocking, and apply the selected control step to the managed control means as a temporary control overlay having an expiration time and an automatic return condition without overwriting the stored standard policy.
[0029] The above one or more processors can distinguish and acquire from one another the adaptive behavior of the automation subject, including at least one of a requested control state, an actual applied control state, a control effect identified from subsequent flow metadata of the attribution unit, and a change in the retry interval distribution identified from flow metadata after the application of the temporary control overlay, a switch to an alternative destination, protocol or port, and a distribution to another attribution unit.
[0030] If the adaptive behavior is determined to be a transition or dispersion that evades control, the above one or more processors may raise the control level or extend the scope of application of the temporary control overlay to the attribution unit associated with the adaptive behavior, and, depending on the verification result of the control effect, lower or release the control level or promote the automation subject candidate branch to a new approval standard twin version and a new contract version.
[0031] In one embodiment, the one or more processors input the behavioral features into a learned discrimination model to calculate the automation potential and its reliability, and may store the model's identifier and version, input evidence identifier and feature-specific contribution in conjunction.
[0032] In one embodiment, the one or more processors may use the adaptive behavior to update the automation possibility, and if the automation possibility is uncertain, apply low-impact control as a discriminant test to update the automation possibility in the form of the time of the response.
[0033] In one embodiment, safety limits regarding protection roles, customer maximum levels, simultaneous isolation upper limits, and enforcement capabilities are applied to the selected control level, said safety limits only lower the control level and not raise it, and when the upper limit is reached, instead of rejecting the request, it may lower it to a control level one level below.
[0034] A method for controlling the behavior of an automated subject network according to another aspect of the present invention is performed by one or more processors that execute program instructions stored in memory, and includes the step of each processing of the system being performed by said one or more processors. According to yet another aspect of the present invention, a non-transient computer-readable recording medium is provided on which a program for executing said method on a computer is recorded. Effects of the invention
[0035] According to the present invention, since the automation subject is identified solely by flow metadata without viewing the communication content, the burden of protecting personal information and communication secrets is reduced, and it is possible to respond to encrypted communication and new services not in the list.
[0036] According to the present invention, by distinguishing between approved and unapproved automation through network behavior contracts by automation subject, repeated false positives of approved business automation can be reduced, and by separating contract deviations into candidate branches, the approval criteria can be prevented from being contaminated.
[0037] According to the present invention, since a temporary overlay is applied starting from a control with a small impact on business operations and the safety limit only lowers the control level, the risk of business interruption can be reduced even when automatic response is turned on.
[0038] According to the present invention, since the request status, actual application status, control effect, and adaptive behavior are distinguished and verified, false success of response commands and bypassing of automation agents can be detected, and the adaptive behavior itself can be used as evidence of automation.
[0039] According to the present invention, device-specific attribution, identification, and control can be performed using only a single edge device acting as a gateway, making it applicable even to small organizations without dedicated security personnel, and accurate control can be executed with minimal information by separating records for analysis and exact value evidence for execution. Brief explanation of the drawing
[0040] FIG. 1 is an overall configuration diagram of an automated subject network behavior control system according to one embodiment of the present invention. FIG. 2 is a flowchart of an automated subject network behavior control method according to one embodiment of the present invention. Figure 3 shows the data structure of the automation subject node, network behavior contract, and automation subject candidate branch of the approval standard network digital twin. Figure 4 shows the configuration of behavioral feature generation, automated discrimination, and artificial intelligence service destination discrimination. Figure 5 illustrates the process of generating a temporary control overlay by passing through a control stage in order of work impact, a counterfactual evaluation, a limitation on the automation level, and a safety inspector that only lowers the control stage. Figure 6 shows the four-layer verification and discrimination test, execution defect determination and determination results of the request status, actual application status, control effect and adaptive behavior. Figure 7 shows the traffic path and observation path of a gateway-type edge device, the separation of pseudonymous records for analysis and exact value evidence for execution, and an execution structure with separated authority. Figure 8 shows a pre-action closed loop leading to pre-action, detailed reporting, subsequent decision, one-time execution of a signature instruction, and a decision record. Figure 9 shows the handling of disconnection and re-control according to the entity responsible for expiration or release. Specific details for implementing the invention
[0041] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. However, the present invention is not limited to the embodiments described below and may be implemented in various different forms, and the embodiments are provided to ensure that the disclosure of the present invention is complete and to inform those skilled in the art of the scope of the invention.
[0042] The terms used in this specification are for describing embodiments and are not intended to limit the invention. The singular form includes the plural form unless the context clearly indicates otherwise, and "comprising" does not exclude the presence or addition of one or more other components in addition to the mentioned components.
[0043] Each functionally separated component in this specification may be implemented by one or more processors executing program instructions stored in memory, and may be integrated into the same computing device or distributed across edge devices, servers, virtual machines, containers, or managed services. Each component may be implemented in hardware, software, firmware, or a combination thereof, and the program instructions may be stored on a non-transient computer-readable recording medium.
[0044] 1. Definition of Terms
[0045] An 'automation subject' refers to an execution entity that generates network communications repeatedly or sequentially without individual human intervention. An automation subject may include large-scale language model-based artificial intelligence agents, agent tool-calling clients, browser extensions, scripts, scheduled tasks, business automation tools, synchronization and backup clients, and external automated reconnaissance and penetration tools, and may include communications that are initiated by human instructions and continue autonomously.
[0046] The 'attribution unit' is an identification unit that attributes a flow to an automation subject, and may be at least one or a combination thereof of a terminal identifier, an executable file hash, signer, or path, a process lineage, a service, container, or workload identifier, a browser extension identifier, a pseudonymized user or role identifier, a TLS client fingerprint, and a hash of an authentication token identifier.
[0047] A 'Network Behavior Contract' is data that represents the scope of authorized network behavior for each automation subject in a way that allows machines to verify it. A Network Behavior Contract may include at least some of a contract identifier, attribution conditions, allowed destinations, allowed internal resources and data classes, allowed data movement directions, speed sheaths, allowed time zones, allowed protocols, ports, and authentication methods, approval policies upon exceeding limits, contract version, validity period, issuer, and digital signature. Allowed destinations may be represented as at least one of a domain, address range, service category, and autonomous system.
[0048] The 'rate cap' is a set of upper or lower limits for the number of requests per time window, the number of concurrent connections, transmission volume, minimum call interval, and continuous activity time. The 'human rate cap' is the default rate cap applied when there is no rate cap in the network behavior contract, representing the range of request frequency, minimum call interval, and continuous activity time achievable through individual human operations. The 'rate cap' is a value that prevents communication within an assigned unit from being exceeded during deceleration control, and is determined from the rate cap or the human rate cap.
[0049] 'Behavioral features' refer to temporal features, structural features, size features, and fingerprint features derived from flow metadata without decoding the communication content. 'Automaticity' is a value and its reliability indicating the probability that a communication of an attribution unit was generated without individual human manipulation.
[0050] "Contract deviation" refers to a state in which observed behavior falls outside the scope of the network behavior contract and may include unregistered principals, destination deviations, data class deviations, data movement direction deviations, velocity deviations, time zone deviations, and behavioral pattern deviations. An unregistered principal refers to an automated principal without a network behavior contract, and a behavioral pattern deviation refers to a case where the pattern changes abruptly compared to an approved behavioral profile, even if the communication is within the contract scope.
[0051] The 'Acceptance Baseline Network Digital Twin' is a twin that represents the verified and accepted current state along with its version. The Acceptance Baseline Network Digital Twin can include assets, pseudonymized identities, services, data resources and data classes, automation principals, network behavior contracts, communication relationships, data movement relationships, and policies as nodes, edges, or attributes, and can associate a source, time of observation, trust level, and version with each state. The 'Automation Principal Candidate Branch' is a candidate state that holds the base version identifier and change delta of the Acceptance Baseline Network Digital Twin and is managed without overwriting the Acceptance Baseline Network Digital Twin.
[0052] 'Control stages' are stages of control arranged in order of business impact, and may be arranged in the order of, for example, maintaining observation, strengthening observation, deceleration within a speed limit, concurrency or session restriction, destination category restriction, confirmation request, destination blocking, and attribution unit isolation. Depending on the embodiment, a warning stage that provides only notification may be included. 'Temporary control overlay' is a control applied as a higher priority or separate policy layer without deleting or overwriting the stored baseline policy, and has a target attribution unit, control stage, parameters, baseline policy version, expiration time, and automatic revert condition.
[0053] 'Adaptive behavior' refers to behavioral changes exhibited by the automation agent after the application of control, such as changes in retry intervals, switching to alternative destinations, protocols, or ports, and distribution to different attribution units.
[0054] 2. System Configuration
[0055] Referring to FIG. 1, a system according to one embodiment of the present invention may include an edge device (100) positioned between a managed network (10) and an external network (20), a central server (200) that exchanges signed reports and instructions (260) with the edge device (100), and an operator terminal (300).
[0056] The managed network (10) includes terminals (12) and internal server / data resources (16), and automation entities (14), such as artificial intelligence agents, scripts, and browser extensions, can be executed on the terminals (12). Data grades may be assigned to the internal server / data resources (16). The external network (20) may include artificial intelligence service destinations (22), other external destinations (24), and external automation entities (26).
[0057] The edge device (100) includes a memory (102) and a processor (104), and by the processor (104) executing program commands stored in the memory (102), it can implement an observation information integration unit (110), an attribution / subject management unit (120), a behavioral feature calculation unit (130), an automated discrimination unit (140), a contract comparison unit (150), a stepwise control unit (160), an application / effect verification unit (170), and a personal information protection unit (180). The edge device (100) may include an execution unit (190) with separated authority and a control means (195), such as a packet filter table.
[0058] The central server (200) includes memory (202) and a processor (204) and can implement a versioned twin management unit (210), a contract management unit (220), a candidate branch management unit (230), a promotion / return management unit (240), and an explanation / report unit (250). The versioned twin management unit (210) manages an approval standard network digital twin (212), the contract management unit (220) manages a network behavior contract (222), and the candidate branch management unit (230) manages an automation subject candidate branch (232).
[0059] The observation information integration unit (110) receives flow metadata from gateways, routers, switch mirror ports, firewalls, proxies, DNS logs, flow records, and host sensors. The flow metadata may include time, 5-tuples, number of bytes and packets per direction, connection duration, DNS query and response names, TLS server names, application layer protocol negotiation values, client fingerprints, certificate metadata, retransmissions and resets, and a summary of record size and arrival intervals. The observation information integration unit (110) does not collect payloads by default.
[0060] The attribution and subject management unit (120) groups flows into attribution units to create and update automated subject nodes and assign attribution reliability. If a host sensor is present, the attribution and subject management unit (120) can use the executable file hash, signer, process lineage, and user input idle time without input content together.
[0061] The behavioral characteristic calculation unit (130) calculates behavioral characteristics by attribution unit and time window, and the automation determination unit (140) calculates automation possibility, reliability, and basis characteristics using rules, learned determination models, or a combination of both. The contract comparison unit (150) compares the observed behavior with the network behavior contract (222) to determine the type and severity of contract deviation.
[0062] The stepwise control unit (160) generates candidate control stages and evaluates them counterfactually to select one, and applies the selected control stage to the control means (195) through the execution unit (190) as a temporary control overlay. The application and effect verification unit (170) distinguishes and obtains the requested control state, the actual application state, the control effect, and the adaptive behavior.
[0063] The promotion and return management department (240) raises, lowers, or releases the control level based on the verification results, and promotes the automation subject candidate branch (232) to a new approval standard twin version and a new contract version, or returns it to the latest approval standard. The explanation and report department (250) creates sentences with evidence grades based on the grounds for judgment and control and generates period-based reports.
[0064] The personal information protection department (180) performs identity pseudonymization, phased identity de-identification, edge-local preservation of the original, and upper-level transmission of only events and aggregations. The operator terminal (300) performs contract approval, subsequent decision, and report inquiry through the central server (200).
[0065] The configuration arrangement of FIG. 1 is an example. For instance, a copy of the approval standard network digital twin (212) and a candidate branch management function may be placed in the edge device (100), and observation, feature calculation, identification, and control application may be performed in the edge device (100), while contract approval, reporting, and multiple site management may be performed in the central server (200). Even if the connection with the central server (200) is lost, the edge device (100) independently performs the expiration and automatic return of the temporary control overlay.
[0066] 3. Processing Flow
[0067] Referring to FIG. 2, in step S210, one or more processors receive flow metadata that has not been decoded from one or more network observers. In step S220, one or more processors group the flow metadata into attribution units to form candidates for automation subjects.
[0068] In step S230, one or more processors calculate behavioral characteristics for each belonging unit, including at least two of call interval distribution, clustering, fan-out, transmission and reception ratio, and response delay, and determine the possibility of automation to identify the automation subject. The identified automation subject is managed as a node of the approval standard network digital twin (212), or, if not yet approved, is managed as a candidate.
[0069] In step S240, one or more processors compare the observed behavioral features, destinations, and data movements with the network behavior contract (222). If there is no contract deviation, the observation continues, and if there is a contract deviation, the process proceeds to step S250.
[0070] In step S250, one or more processors do not directly merge the contract exit into the approval reference network digital twin (212) but instead generate an automated subject candidate branch (232) containing a reference version identifier and a change delta, and set a situation hypothesis and an evidence completion condition.
[0071] In step S260, one or more processors generate control level candidates, evaluate them counterfactually, and select a control level through a safety check that only lowers the control level. In step S270, one or more processors apply the selected control level as a temporary control overlay having an expiration time and an auto-return condition without overwriting the stored reference policy.
[0072] In step S280, one or more processors distinguish and obtain the requested control status, actual application status, control effect, and adaptive action. Based on the verification results, if confirmed as approved automation, it is promoted to a new version of the contract and twin (S291); if risk persists or bypass is confirmed, the control level is raised or the scope of application is expanded (S292); if effect is confirmed and risk is resolved, the control level is lowered or released and it returns to the latest approval standard (S293); and if there is insufficient evidence, follow-up observation is performed (S294).
[0073] 4. Data Structures
[0074] Referring to FIG. 3, the approval standard network digital twin (212) may include terminal / asset nodes (12), data resource nodes (16), automation subject nodes (14), and destination nodes (22, 24). The terminal / asset nodes and automation subject nodes may be connected by an execution relationship, the data resource nodes and automation subject nodes by a data movement relationship, and the automation subject nodes and destination nodes by a communication relationship. The destination nodes may be represented by a service category or address range reference.
[0075] The automation subject node (14) may have attributes such as a subject identifier, an attribution unit, automation capability and trust, an authorized behavior profile, and a contract reference. The attribution unit may include at least one of a terminal identifier, an execution hash / signer, a pseudonym user / role, and a TLS client fingerprint.
[0076] A network behavior contract (222) may include a contract identifier, contract version, validity period, issuer signature, attribution condition, allowed destination, allowed internal resource / data class / data movement direction, speed shell, allowed time zone, allowed protocol / port / authentication method, approval policy upon exceeding, and contract source.
[0077] Contract sources may include administrator registration, transformations from the agent execution system's manifest, and administrator approval of contract drafts generated from observer profiles. Transformation from the manifest involves converting the execution providers and destinations listed in the agent execution system's tool allowlist, allowed tools and blocked tools, approval policies, or function contract into the allowed destinations and approval policies of the network behavior contract. Contract drafts are generated from the observed behavior of automation principals identified as approved business automation and are registered as new contract versions after undergoing administrator approval.
[0078] Contracts can be hierarchized by organization, department, and subject, and lower-level contracts cannot extend the scope of higher-level contracts. One or more processors may use only contracts for which the issuer's electronic signature and validity period have been verified to determine contract withdrawal.
[0079] The automation subject candidate branch (232) may include a branch identifier, status, reference twin version identifier, change delta, contract breakout type, situation hypothesis, evidence completion condition, and unfulfilled evidence. The automation subject candidate branch (232) is bound to the reference twin version and is not directly merged into the approval criteria.
[0080] 5. Behavioral Characteristics and Automation Identification
[0081] Referring to FIG. 4, the behavioral feature calculation unit (130) receives a window summary record (116) and optionally host sensor information. The window summary record (116) may include, for example, the first and last observation times per flow, the number of bytes and packets per direction, and pseudonym references for every 60-second window, and the host sensor information may include executioner attribution and input idle time with no input content.
[0082] Time characteristics (131) may include the mean of the call interval, variance, coefficient of variation, entropy, discreteness of the interval distribution, periodicity, 24-hour activity distribution, continuous activity time without interruption, cluster index, response delay, and transmission during input idle. The cluster index may be the value obtained by dividing the number of requests in a short time window by the long-term average. The response delay is the time from the end of receiving the response to the start of the next request, and if this time is repeated shorter than a predetermined standard reflecting human reading and input time, it is used as evidence of automation.
[0083] The structural features (132) may include a unique external destination within the time window, a unique internal host and a number of unique ports, a fan-out, a sequential or response-based adaptive search order of addresses, ports, and paths, a sequence such as external transmission after internal resource access, a number of simultaneous connections, and a retry type such as exponential backoff or a fixed interval.
[0084] Size features (133) may include a transmission / reception byte ratio, a degree of upload dominance where the amount transmitted is greater than the amount received, a request size distribution, and a streaming response form. The streaming response form is a form in which a received record of less than or equal to a predetermined size within a long-term connection is sustained for a predetermined interval range, and is used to determine whether it matches the token unit response of the generative model.
[0085] Fingerprint features (134) may include whether the TLS client fingerprint matches the associated executor or claimed client, the non-browser client library fingerprint, and the simultaneous appearance of multiple terminals of the same fingerprint.
[0086] In one embodiment, behavioral characteristics can be calculated from a window-level summary without preserving packet-level times. For example, the call interval is approximated from the window-level first observation time of flows heading to the same destination from the same attribution unit, crowding and fan-out are calculated from the number of flows, new relationships, and different remote references per window, and the transmit / receive ratio can be calculated from the window-level directional bytes.
[0087] The automation determination unit (140) produces a determination result (143) including automation possibility, reliability, and feature contribution by using a rule score (141) and a learning determination model (142) alone or in combination. The learning determination model (142) may be a gradient boosting model, a time series model, or a graph model, and may be trained by using the flow of automation agents registered in the network behavior contract (222) as positive, the flow where human interaction is confirmed as negative, and the operator's judgment as a correction label. The determination result (143) is linked with a model identifier and version, an input evidence identifier, and feature contribution, and is used for auditing and explanation.
[0088] Human decisions are stored as training labels, while AI recommendations and external model judgments are stored as separate criteria to prevent them from mixing with the correct labels. One or more processors can determine whether to update the discrimination model by measuring the degree of agreement between recommendations and decisions.
[0089] Artificial intelligence service destination determination (145) assigns a category to destinations that match a known list of artificial intelligence service destinations, such as domains, TLS server names, and address ranges. Even if a destination is not included in the list, if a streaming response type is observed and at least one of upload dominance and application interface type connection patterns is observed together, that destination is marked as a generative artificial intelligence candidate destination and managed as a candidate until confirmation by an operator. The determination result is provided to the contract comparison unit (150).
[0090] 6. Contract Comparison and Candidate Branching
[0091] The contract comparison unit (150) calculates the type and severity of contract deviation by comparing each of the destination, data movement, speed, time zone, and behavioral pattern with the contract items.
[0092] Data class deviation can be determined in a chain. If the same belonging unit accesses an internal resource having a data class higher than the reference class within a predetermined time window, and then the amount of transmission to an external destination not allowed by the network behavior contract increases, one or more processors determine that there is a data class deviation and include candidate edges and outflow path hypotheses regarding data movement relationships in the automation subject candidate branch (232).
[0093] A deviation in behavioral pattern occurs when, even within the scope of the contract, at least one of the automation principal's fan-out, search order, and response delay changes by more than a predetermined threshold compared to an approved behavioral profile. In this case, one or more processors generate a situational hypothesis regarding the automation principal that has been hijacked or whose behavior has been altered by instructions in a document or tool response.
[0094] Situational hypotheses may include authorized business automation, unauthorized agents installed by users, agents with altered behavior, external automation intrusion, high-speed human operations, or false positives. The evidence completion condition consists of evidence items and satisfaction criteria per hypothesis, and evidence items may include verification of the executor signature, verification of the user or administrator, verification of the destination category, verification of the data class, and alignment with the change schedule.
[0095] If there is unsatisfactory evidence, one or more processors plan a follow-up observation. Follow-up observations may include host sensor attribution requests, time window expansion, DNS resolution history lookups, and flow summary resolution upgrades; upgrades to the level of viewing communications are permitted only when separate authorization is confirmed.
[0096] Relationships not included in the approval criteria may be graded based on the degree of deviation. For example, authorized outgoing relationships not in approved flow profiles are classified as Record grade; if transmissions to new relationships exceed the lower limit of the consecutive window total, they are classified as Warning grade; if events within the same set of relationships repeat across multiple windows or multiple new relationships appear in a single window, the classification is raised to Warning; and repetitions or surges originating within a maintenance window declared by the owner may be downgraded to Record grade. Relationships and names verified by the administrator are reflected in the allowlist, but the approved flow profiles themselves are not modified.
[0097] 7. Selection of Control Stages and Safety Inspection
[0098] Referring to FIG. 5, the phased control unit (160) generates control stage candidates among control stages (161) based on the type and severity of contract deviation, automation possibility, and data grade. The control stages (161) may be arranged in order of business impact, for example, maintaining observation, strengthening observation, decelerating to within the speed limit, concurrency / session limit, destination category limit, verification requirement, destination blocking, and attribution unit isolation.
[0099] Counterfactual evaluation (162) evaluates each candidate in the counterfactual twin branch of the approval standard network digital twin (212) for risk reduction, operational impact, observability maintenance, return costs, and authorization satisfaction. Operational impact may include whether the approved automation is interrupted and the impact on dependent services. The automation levels of the operational policy, such as observation-only, recommendation-only, approval-based execution, restricted execution, and emergency automatic defense, limit the actionable steps.
[0100] The safety checker (163) can be implemented as a pure function that only lowers the selected control level and does not raise it. The safety checker (163) checks, for example, in the following order: if there is a bad suggestion or no valid policy, it is lowered to observation; if there is a request higher than the policy level, it is lowered to the policy level; if there is a match with a policy exception, or if there is a site-wide scope or a central connection path, it is lowered to warning; and if there is a protection role such as the edge device itself, a router, a DHCP server, or a DNS server, it is lowered to warning.
[0101] Next, the safety inspector (163) lowers the level to warning if there is no gateway enforcement path or if the survival signal of the enforcement unit is longer than a predetermined time, lowers the level to that level if it exceeds the maximum level set by the customer, lowers the level to block if it reaches the simultaneous isolation upper limit, and lowers the level to deceleration if it is a block of the entire device range. Above, warning and observation are examples of control levels that do not involve network enforcement.
[0102] Requests that hit the cap are not rejected but are lowered one level to maintain control. In non-enforceable deployments, such as host-only deployments, the policy itself is restricted to a level below which network enforcement does not occur.
[0103] The selected control stage (164) is converted into a temporary control overlay (166). The temporary control overlay (166) has a target attribution unit, stage, parameters, a reference policy version, an expiration time, and an automatic return condition, and upon return, instead of the policy at the time the control started, it retrieves and applies the latest approval reference policy.
[0104] Deceleration is applied using at least one of the following: attribution unit, destination-specific token bucket, minimum call interval, concurrent connection limit, and transmission volume limit; the rate limit is determined from the rate envelope of the network behavior contract or the human rate envelope. Management channels and evidence collection channels are excluded from deceleration.
[0105] A confirmation request requests a response from a terminal user or administrator by presenting at least one of the following without communication content: the alias or executor name of the automation principal, the destination category, the transmission volume, the relevant data class, and the duration. If no response is received within a predetermined timeframe, default controls defined in the policy, such as deceleration or blocking, are maintained. Blocking and isolation are applied while maintaining the management path, return path, and status reply path.
[0106] The control adapter converts the temporary control overlay (166) into commands for actual control means, such as packet filter rules, traffic shaping, firewalls, proxies, DNS response policies, process-unit network restrictions of endpoints, and range restrictions of authentication tokens.
[0107] 8. Four Layers of Validation and Adaptive Behavior
[0108] Referring to FIG. 6, when a temporary control overlay (166) is applied, the application / effect verification unit (170) distinguishes and obtains the requested control state (171), the actual application state (172), the control effect (173), and the adaptive behavior (174).
[0109] The actual application status (172) can be obtained from the success or failure of the table replacement reported by the execution unit (190) and the set of operations during application, and can be verified by rereading the rule applied to the control means (195) and comparing it with the desired status.
[0110] The control effect (173) is confirmed in the subsequent flow of the belonging unit by the transmission speed, destination and whether the data movement has entered the target range, the rule-specific counter, and the suspected leak calculated by an independent observer. For example, if the same target's response byte is seen on the sensor even after a predetermined time has passed since the block was applied, it is determined to be a suspected leak.
[0111] Adaptive behavior (174) includes changes in the distribution of retry intervals after control, switching to alternative destinations, protocols, or ports, dispersion to other affiliated units, and blocked attempts. The decision unit (175) uses the adaptive behavior to update the automation possibility, and if a switching or dispersion that evades control is identified, it determines it as a bypass attempt to raise the control level or extend the scope of application of the temporary control overlay to the relevant affiliated units.
[0112] The adjudication unit (175) also incorporates the automated evidence into the evidence completion conditions, lowers or releases the control level to return to the latest standards when the effect is confirmed and the risk is resolved, and promotes it to a new version of the contract and twin when it is confirmed as approved automation. Verified controls and secured evidence are preserved, and only failed or incomplete controls are replanned.
[0113] The determination test (176) applies at least one low-impact control among delay, pause, and deceleration at a predetermined minimum range and short expiration time when the possibility of automation is in a predetermined uncertainty range, and compares the time pattern of the response of the attribution unit to it. Since human operation responses are irregular and have long delays, whereas automated retries tend to take the form of fixed-interval retries or exponential backoff, one or more processors update the possibility of automation through this comparison.
[0114] Execution defect determination (177) is based on comparing the survival signal of the execution unit with the report of the edge, and if the work held by the execution unit is more than the report, or if there is no survival signal, or if the table replacement failed, or if there is a surge in requests, it is raised as an execution defect.
[0115] 9. Example of a gateway-type edge device
[0116] Referring to FIG. 7, the edge device (100) of the present embodiment has a single network interface, a LAN interface (106), and is converted into a gateway so that traffic from the terminal (12) goes out to an external network (20) through the LAN interface (106), a logging hook (112), a packet filter table (195), and an address translation (108).
[0117] The logging hook (112) is placed at a stage prior to address translation (108) and prior to the control rule of the packet filter table (195). Thus, the sensor (114) can observe the internal terminal address before address translation to obtain attribution for each internal terminal, and can continue to observe attempts blocked by the control rule. The blocked attempts are used to determine the control effect and the action after control.
[0118] The sensor (114) reads only the packet header with a short capture length, e.g. 96 bytes, and does not preserve the payload. The sensor (114) summarizes the first and last observation times per flow, the number of bytes and packets per direction, TCP flags, connection status, and direction estimation for each window, signs the summary with a device key, and passes it to an authorized runtime (119). The runtime (119) records only after verifying the signature, owner approval, and authorization policy.
[0119] In the pseudonymized record (116) for analysis, the address and port are converted into a key hash reference using a site-specific secret key and stored as a flow, remote counterpart, remote service, remote address range, e.g. IPv4 / 24 and IPv6 / 48, and range / service reference, and the original address is not left in the analysis record. The remote address range reference can serve as a unit for grouping destinations without a destination name and can be used as a substitute for destination categories for reference learning and control ranges.
[0120] The DNS name can be obtained only when there is owner approval in the response directly resolved by the edge device (100) and attached to the remote counterpart's key hash reference. When sending to a server outside the edge device (100), the destination domain name can be transmitted only after the edge device (100) converts it into a destination category code or key hash, so that the name itself is not sent out.
[0121] The sensor (114) separately loads the exact device and destination address and port as execution accuracy value evidence (118) within a predetermined number of cases per window, only for flows that are candidates for control, such as flows going to a risk service or flows where the window transmission volume is above a lower limit. The runtime (119) recalculates the key hash reference with the accuracy value to confirm the binding, uses the evidence only for the control plan, and leaves only the key hash reference in the analysis record and the upper report. After the control ends and a predetermined period, such as 30 days, passes, the accuracy value is deleted.
[0122] When the runtime (119) generates a control request through determination, contract comparison, and control planning, the execution process of the execution unit (190), which is separated from the runtime process and authority, atomically replaces the packet filter table (195) only after re-examining the request against its fixed limits. The fixed limits may include, for example, a limit on the number of operations, a limit on the lifetime, a time error, checking the size, owner, and link of the request file, a unicast destination range outside the subnet, and a protection address denial.
[0123] The enforcement unit (190) may apply an upper limit on the number of controlled objects proportional to the site size N. For example, the number of devices to be touched may be limited to the value rounded up from 50% of N, the number of quarantined may be limited to the value smaller between 3 and the value rounded up from 20% of N, and 20 may be limited, and a surge blocking measure may be implemented to reject all new requests when the number of rejections per day reaches a lower limit.
[0124] In the packet filter table (195), deceleration may include a form that slows down the flow without interrupting by discarding only the portion exceeding the speed limit during transmission from the device to the destination or between the destination and the service, e.g., 16 kilobytes per second and 64 kilobytes per burst, and a device unit form that discards the portion exceeding the speed limit during new connections opened by the device outside the subnet, e.g., 2 connections per second and 20 bursts. Blocking is executed at the device-to-destination, protocol, and port unit, and isolation is executed at the bidirectional unit between the device and outside the subnet.
[0125] Destination list control blocks by address range set and leaves attempted device-destination pairs in a dynamic set with expired, for example, one hour, to serve as evidence for device-specific warnings and adaptive actions. In the same way, blocking or slowing can be applied to address range sets by destination category.
[0126] In the learning of the approval criteria, phase baseline candidates and flow profile candidates must meet the number of observations and observation intervals, e.g., phase 2 times and 3,600 seconds, and flow 3 times and 86,400 seconds, to be awaiting approval. Approval is performed by a command bound to a candidate identifier, digest, and idempotent key, and only a set of relationships of a named plate is approved. If the policy allows, candidates that meet the learning conditions can be conditionally automatically approved. The control policy bundle can be distributed by negotiating based on plate information carried in the survival signal of the edge device (100).
[0127] 10. Preemptive Closed Loop
[0128] Referring to FIG. 8, the edge device (100) first performs a preliminary action of applying a temporary control overlay within the policy scope (S810), and sends a detailed report including a key hash reference, step, reason, and application status to the central server (200) (S820).
[0129] The central server (200) receives the report and records it as an event (S830), and when a subsequent decision is made at the operator terminal (300), such as release, retention, or time-limited user verification request (S840), it generates a signature instruction having an instruction identifier, sequence number, expiration, and idempotent key (S845).
[0130] The edge device (100) verifies the signature instruction and checks the idempotent key, sequence number, and expiration, executes it only once (S850), replies the execution result (S860), and the central server (200) leaves a decision record (S870).
[0131] 11. Disconnection and Re-control
[0132] Referring to FIG. 9, the edge device (100) monitors the connection status with the central server (200) as online, retry, and offline (S910). Controls imposed while the connection is disconnected are maintained for their respective lifespans (S920), and a decision to extend or release them can be made on-site by stating the reason, and the decision is recorded (S930). When reconnecting, the on-site decision is reported to the central server (200) and reflected in the decision record (S940).
[0133] When control expires or is released (S950), one or more processors determine the timing of re-control differently depending on the release agent. Since expiration is not an exemption, control is re-implemented if the same behavior continues (S961). If released by a decision of the central server, the same target is not re-implemented for a predetermined exemption period (S962). If released by an emergency release at the site, control is re-implemented if the same behavior is observed after a waiting period shorter than the exemption period (S963).
[0134] 12. Examples
[0135] Example 1 relates to the export of data by an artificial intelligence agent installed by a user. After the agent of the office terminal (12) accessed a file server with a data grade of 'Internal Confidential', the amount of transmission to the external language model application program interface increased to tens of times the usual amount within 5 minutes.
[0136] One or more processors generate an automated subject candidate branch (232) with an unregistered subject and a data class cascade out of the way, and place the executor signature, user verification, and destination category as unsatisfactory evidence. In the counterfactual assessment, the deceleration has the greatest risk reduction relative to business impact, so a 15-minute overlay with a transmission limit and a minimum call interval and a verification requirement are applied.
[0137] When an agent switches to a different AI destination during effectiveness verification, one or more processors determine this to be a bypass attempt and escalate the level to a destination category restriction. If the administrator determines it to be unauthorized, the block is maintained, and only the event and reasoning flow references are preserved without the communication content.
[0138] Example 2 concerns the avoidance of false positives and contract renewal of approved business automation. When a nightly backup client begins writing to a new cloud storage destination, one or more processors keep the destination exit as a candidate branch but do not include it in the approval criteria, and remain in the observation maintenance phase. When the evidence completion condition is met with a change schedule and administrator confirmation, a draft contract is created with the observation profile, and upon approval, it is promoted to a new contract version.
[0139] Example 3 relates to an AI service not on the list. When an unlisted destination appears together with upload dominance, streaming response type, and non-browser client fingerprint, that destination is marked as a candidate destination for generative AI and listed as an item requiring verification in the report. When classified by an administrator, the destination category is updated.
[0140] Example 4 relates to a change in the behavior of a registered agent. When a business agent contracted to use only the internal wiki, ticket system, and specified language model drastically increases internal host fan-out and port exploration within the scope of the contracted destination, one or more processors make a hypothesis about the agent whose behavior has changed due to a deviation in behavioral pattern, apply internal destination category restrictions and administrator verification, and then verify the effect.
[0141] Example 5 relates to external automation intrusion. When a request entering a remote access / web service shows adaptive discovery that changes the path and parameters according to the response to a response shorter than that of a human, one or more processors create candidates for external automation entities (26) and apply delayed responses and deceleration. When a change in the retry interval corresponding to the delay is observed, it is reflected as automation evidence and raised to blocking, and an evidence package can be delivered to the counterparty business operator within the scope of verified authority.
[0142] Example 6 relates to attribution by deployment type. Small organizations use only one router or gateway and use a combination of a terminal identifier and a TLS client fingerprint as the attribution unit, while organizations with host sensors use an executable file hash, signer, process lineage, and input idle time together. Managed security service providers operate twins and contracts separately for each customer.
[0143] Example 7 relates to a closed network and a public institution. A learning discrimination model (142) is executed in a computational area separated from the external network, and only events and aggregations are sent to the upper system, and model updates are imported only through packages with verified signatures.
[0144] Example 8 relates to a period report for small organizations. The explanation and report section (250) aggregates the number of automated subjects, the trend of transmission volume by artificial intelligence destination, unregistered subjects, contract deviations, and control history on a monthly basis, attaches one of the evidence grades (confirmed, likely, possible, unconfirmed, and refutation) and a reference to the basis flow to each sentence, and does not include communication content.
[0145] 13. Variations
[0146] Observers can be extended to kernel-level packet observations, operating system event traces, proxy logs, DNS resolver logs, cloud virtual network flow logs, and security access service logs. Units of attribution can be extended to service accounts of containers or orchestrators, workload identities, hashes of application interface keys, and identifiers of remote tool servers.
[0147] Contracts can be represented as structured schemas, policy languages, or signed manifests. If an agent sends a request containing a contract identifier or a signed identity, this can be used as the basis for attribution.
[0148] Discrimination can be composed of two stages: an edge lightweight model and a central model, and federated learning using feature statistics from multiple sites can be applied. The language model that generates explanations only proposes explanation and control candidates, and only candidates that pass deterministic checks of schema, authority, and invariant conditions are selected for execution.
[0149] Control measures can be extended to the policy application interfaces of security access services, process-level network blocking of endpoint security products, and token range reduction of authentication providers. The arrangement of control steps may be reconfigured according to organizational policy, or some steps may be omitted.
[0150] The system provides contract and verification results to the upper AI control plane and can jointly evaluate the impact of changes in other operational domains on the contracts of the automated principal.
[0151] Although embodiments of the present invention have been described above, the present invention is not limited thereto and may be implemented with various modifications within the scope of the technical concept described in the claims. The numerical values described in the embodiments are examples and may be changed according to policy. Industrial applicability
[0152] The present invention can be applied to network security devices, managed security services, security access services, and artificial intelligence agent operating environments of enterprises, public institutions, and small business establishments, and can be used to provide stepwise control that identifies the communications of automated entities without viewing their content and minimizes business impact. Explanation of the symbols
[0153] 10: Managed Network, 12: Terminal, 14: Automation Subject, 16: Internal Server / Data Resource, 20: External Network, 22: AI Service Destination, 24: Other External Destination, 26: External Automation Subject, 100: Edge Device, 102, 202: Memory, 104, 204: Processor, 106: LAN Interface, 108: Address Translation, 110: Observation Information Integration Unit, 112: Logging Hook, 114: Sensor, 116: Pseudonymous Record for Analysis (Window Summary Record), 118: Executional Exact Value Evidence, 119: Runtime, 120: Attribution / Subject Management Unit, 130: Behavioral Feature Calculation Unit, 131: Temporal Feature, 132: Structural Feature, 133: Size Feature, 134: Fingerprint Feature, 140: Automation Identification Unit, 141: Rule Score, 142: Learning Discrimination Model, 143: Discrimination Result, 145: AI Service Destination Discrimination, 150: Contract Comparison Unit, 160: Phased Control Unit, 161: Control Phase, 162: Counterfactual Evaluation, 163: Safety Checker, 164: Selected Control Phase, 166: Temporary Control Overlay, 170: Application / Effect Verification Unit, 171: Requested Control State, 172: Actual Application State, 173: Control Effect, 174: Adaptive Behavior, 175: Judgment Unit, 176: Discrimination Test, 177: Execution Fault Judgment, 180: Privacy Protection Unit, 190: Execution Unit, 195: Control Means (Packet Filter Table), 200: Central Server, 210: Versioned Twin Management Unit, 212: Approval Criteria Network Digital Twin, 220: Contract Management Unit, 222: Network Behavior Contract, 230: Candidate Branch Management Department, 232: Automation Subject Candidate Branch, 240: Promotion / Return Management Department, 250: Explanation / Report Department, 260: Signed Reports / Instructions, 300: Operator Terminal
Claims
Claim 1 Memory for storing program instructions; An automated subject network behavior control system comprising one or more processors executing the above program commands, wherein the one or more processors receive flow metadata with undecoded communication content from one or more network observers, group the flow metadata into at least one attribution unit including a terminal identifier, an attribution entity identifier, a pseudonymous user identifier, and a client fingerprint, calculate behavioral features for each attribution unit including at least two of call interval distribution, clustering, fan-out, transmission / reception ratio, and response delay, and identify one or more automated subjects by determining an automation possibility indicating the likelihood that the communication of the attribution unit was generated without individual human operation based on the behavioral features, and manage an approval standard network digital twin including the automated subjects as nodes, and a network behavior contract for each automated subject including at least one of an allowed destination, an allowed data class, a rate envelope, and an allowed time zone, and a contract version, and determine at least one contract deviation among an unregistered subject, a destination deviation, a data class deviation, a rate deviation, and a time zone deviation by comparing the observed behavioral features, destination, and data movement with the network behavior contract, and without directly merging the contract deviation into the approval standard network digital twin, the Creates an automation subject candidate branch including a reference version identifier and a change delta of the approval reference network digital twin, and based on the type of contract deviation and the automation possibility, selects one of a plurality of control steps arranged in order of business impact, including at least three of maintaining observation, deceleration within a rate cap, concurrency or session limit, destination category limit, confirmation request, and blocking, and applies the selected control step to a managed control means as a temporary control overlay having an expiration time and an auto-return condition without overwriting the stored reference policy, and the requested control state, the control state actually applied to the managed control means,An automation subject network behavior control system characterized by distinguishing and acquiring from one another the control effect identified from the subsequent flow metadata of the above-mentioned attribution unit, and the adaptive behavior of the above-mentioned automation subject identified from the flow metadata after the application of the above-mentioned temporary control overlay, which includes at least one of a change in the retry interval distribution, a switch to an alternative destination, a protocol or port, and a dispersion to another attribution unit; and if the adaptive behavior is determined to be a switching or dispersion that evades control, the control level is raised or the scope of application of the above-mentioned temporary control overlay is expanded to the attribution unit associated with the adaptive behavior, and the control level is lowered or released or the automation subject candidate branch is promoted to a new approval standard twin version and a new contract version according to the verification result of the control effect. Claim 2 An automation subject network behavior control system according to claim 1, wherein the one or more processors input the behavioral features into a learned discrimination model to calculate the automation possibility and the reliability thereof, and are configured to store the identifier and version of the learned discrimination model, the evidence identifier of the flow metadata, and the contribution of each behavioral feature in conjunction with the automation possibility, and wherein the learned discrimination model is learned from labels obtained from at least one of the flow of an automation subject registered in the network behavior contract, the flow in which human interaction is confirmed, and the judgment of an operator. Claim 3 An automation subject network behavior control system according to claim 1, wherein the network behavior contract is generated by at least one of administrator registration, a tool allowlist of an agent execution system, allowed tools and blocked tools, a conversion from an approval policy or function contract, and administrator approval of a draft contract generated from the observed behavior of an automation subject identified as an approved business automation, and wherein the one or more processors are configured to use the network behavior contract for determining contract deviation only when the issuer's electronic signature and validity period of the network behavior contract have been verified. Claim 4 An automated subject network behavior control system according to claim 1, wherein the one or more processors are configured to determine that there is a data class deviation when the amount of transmission to an external destination not permitted by the network behavior contract increases after the same attribution unit accesses an internal resource having a data class higher than or equal to a reference class within a predetermined time window, and to include candidate edges and outflow path hypotheses regarding the data movement relationship of the approval reference network digital twin in the automated subject candidate branch. Claim 5 An automated subject network behavior control system according to claim 1, wherein the behavioral characteristics include at least one of a coefficient of variation of the call interval, entropy, and periodicity, the degree to which the time from the end of receiving a response to the start of the next request is repeated shorter than a predetermined standard, and at least one of user input idle time provided by a host sensor without input content and the simultaneous occurrence of transmission. Claim 6 An automated subject network behavior control system according to claim 1, wherein the one or more processors are configured to mark a destination as a generative AI candidate destination and manage it as a candidate until confirmation by an operator, if a streaming response pattern in which a receiving record of a predetermined size or smaller is observed within a long-term connection for a destination not included in a known list of AI service destinations, and at least one of an upload dominance pattern in which the transmitting amount is greater than the receiving amount and an application programming interface type connection pattern are observed together. Claim 7 An automated subject network behavior control system according to claim 1, wherein, when one or more network observers are configured as gateways, the attribution unit includes a combination of the terminal identifier and the TLS client fingerprint, and when one or more network observers include host sensors, the attribution unit further includes at least one of an executable file hash, a signer, and a process lineage. Claim 8 An automated subject network behavior control system according to claim 1, wherein the one or more processors are configured to select a control step by evaluating each candidate step among the plurality of control steps in the counterfactual twin branch of the approval standard network digital twin as at least one of risk reduction, business impact, maintenance of observability, return cost, and authorization satisfaction. Claim 9 An automated subject network behavior control system according to claim 1, wherein the speed limit is determined from a speed shell included in the network behavior contract, or from a human speed shell representing a range that can be generated by individual human operation when there is no speed shell in the network behavior contract, and the deceleration is applied as at least one of the token buckets per attribution unit and destination, minimum call interval, simultaneous connection limit, and transmission amount limit, and the management channel and evidence collection channel are excluded from the target of the deceleration. Claim 10 An automation subject network behavior control system according to claim 1, characterized in that the verification request requests a response from a terminal user or administrator by presenting at least one of the pseudonym or executor name of the automation subject, destination category, transmission volume, related data class, and period without communication content, and if there is no response within a predetermined time limit, the basic control defined in the policy is maintained. Claim 11 An automation subject network behavior control system according to claim 1, characterized in that one or more processors are configured to use the adaptive behavior for updating the automation possibility. Claim 12 An automation subject network behavior control system according to claim 11, wherein the above one or more processors are configured to apply at least one low-impact control among delay, temporary suspension, and deceleration at a predetermined minimum range and expiration time when the automation possibility is in a predetermined uncertainty interval, and to update the automation possibility by comparing the time form of the response of the attribution unit to the low-impact control with the form of a human operation response and an automated retry. Claim 13 An automated entity network behavior control system according to claim 1, wherein the one or more processors are configured to determine a behavioral pattern deviation if at least one of the fan-out, search order, and response delay of the automated entity changes by more than a predetermined standard when compared with the approved behavioral profile of the automated entity, even if the communication is within the scope of the network behavior contract, and to include a situational hypothesis regarding the automated entity whose behavior has changed in the automated entity candidate branch. Claim 14 An automated subject network behavior control system according to claim 1, wherein the one or more network observers include a logging hook of an edge device operating as a gateway between an internal terminal and an external network, the logging hook is positioned at a stage prior to address translation and preceding the control rule of the managed control means, and the one or more processors configure the attribution unit with the internal terminal-specific flow observed by the logging hook, and are configured to use the attempt blocked by the control rule for at least one of the determination of the control effect and the determination of the adaptive behavior. Claim 15 An automated subject network behavior control system according to claim 1, wherein the one or more processors convert the address and port of the flow metadata into a key hash reference using a site-specific secret key and store it in an analysis record, separately receive the exact address and port as enforcement evidence within a predetermined number of cases only for flows related to the contract deviation, use the enforcement evidence in the creation of the temporary control overlay only when the binding is confirmed by recalculating the key hash reference from the exact address and port, and delete the exact address and port after a predetermined period has elapsed from the termination of the temporary control overlay. Claim 16 An automated subject network behavior control system according to claim 1, wherein the managed target control means includes a packet filter table, and the execution process separated from the runtime process determining control in the one or more processors applies the temporary control overlay by atomically replacing the packet filter table only when the control request generated by the runtime process passes after being re-examined against a fixed limit including at least one of a task number limit, a lifetime limit, a time error, a destination range, a protected address, and a site size. Claim 17 A method for controlling network behavior of an automated subject, performed by one or more processors executing program instructions stored in memory, comprising: a step in which the one or more processors receive flow metadata in which communication content has not been decoded from one or more network observers, and group the flow metadata into at least one attribution unit including a terminal identifier, an attribution entity identifier, a pseudonymous user identifier, and a client fingerprint; a step in which the one or more processors calculate behavioral characteristics for each attribution unit including at least two of call interval distribution, clustering, fan-out, transmission / reception ratio, and response delay, and determine an automation possibility indicating the possibility that the communication of the attribution unit was generated without individual human operation based on the behavioral characteristics to identify one or more automated subjects; a step in which the one or more processors manage an approval standard network digital twin including the automated subject as a node, and a network behavior contract for each automated subject including at least one of an allowed destination, an allowed data class, a rate envelope, and an allowed time zone, and a contract version; and a step in which the one or more processors compare the observed behavioral characteristics, destination, and data movement with the network behavior contract to identify at least one contract deviation among an unregistered subject, a destination deviation, a data class deviation, a rate deviation, and a time zone deviation. A step of determining and generating an automation subject candidate branch including the reference version identifier and change delta of the approval reference network digital twin without directly merging the above contract deviation into the above approval reference network digital twin;A step in which one or more processors, based on the type of contract deviation and the possibility of automation, select one of a plurality of control steps arranged in order of business impact, including at least three of maintaining observation, deceleration within a rate cap, concurrency or session limit, destination category limit, confirmation request, and blocking, and apply the selected control step to a managed control means as a temporary control overlay having an expiration time and an auto-return condition without overwriting a stored reference policy; a step in which one or more processors obtain, by distinguishing from one another, the adaptive behavior of the automation subject, which includes at least one of a change in retry interval distribution, a switch to an alternative destination, a protocol or port, and a distribution to another belonging unit, as identified from the flow metadata after the application of the temporary control overlay; A method for controlling automated subject network behavior, characterized by including the step of, when one or more processors determine that the adaptive behavior is a transition or dispersion that evades control, raising the control level or expanding the scope of application of the temporary control overlay to an attribution unit associated with the adaptive behavior, lowering or releasing the control level according to the verification result of the control effect, or promoting the automated subject candidate branch to a new approval standard twin version and a new contract version. Claim 18 A method for controlling the behavior of an automated subject network according to claim 17, further comprising the step of: one or more processors applying at least one low-impact control among delay, temporary suspension, and deceleration at a predetermined minimum range and expiration time when the automation possibility is in a predetermined uncertainty range, and updating the automation possibility by comparing the time form of the response of the attribution unit to the low-impact control with the form of a human operation response and an automated retry. Claim 19 A method for controlling automated subject network behavior according to claim 17, wherein the above one or more processors further include a step of evaluating each of the candidate steps among the plurality of control steps in the counterfactual twin branch of the approval standard network digital twin as at least one of risk reduction, business impact, maintenance of observability, return cost, and authorization satisfaction, and wherein the control step is selected based on the result of the evaluation. Claim 20 A method for controlling behavior in an automated subject network according to claim 17, further comprising the step of: one or more processors controlling again if the same behavior continues when the temporary control overlay expires; not controlling the same subject again for a predetermined exemption period when the temporary control overlay is released by a decision of a central server; and controlling again if the same behavior is observed after a waiting period shorter than the exemption period when the temporary control overlay is released by an emergency release at the site. Claim 21 A method for controlling automated subject network behavior according to claim 17, further comprising the step of one or more processors converting the address and port of the flow metadata at the edge device into a key hash reference using a site-specific secret key, and converting the destination domain name into a destination category code or key hash at the edge device, and then transmitting it to a server outside the edge device. Claim 22 A non-transient computer-readable recording medium having a program recorded thereon for executing the method of any one of paragraphs 17 through 21 on a computer.
Citation Information
Patent Citations
Network Security System and Method for applyingSecurity Rule for Restricted Condition
KR1020030003593A
System for determining abnormal crawling
KR102909534B1
Methods and system for blocking ticketing macro bots using artificial intelligence-based user behavior pattern analysis
KR103008759B1
Ai-based security policy learning and anomaly detection system for household networks
KR103009930B1
Network agent for generating platform specific network policies
US11252038B2
