An autonomous attack path probing and self-evolution method and system based on metacognition driving

Abstract

The application discloses a meta-cognition-driven autonomous attack path detection and self-evolution method and system, relates to the field of network security, and comprises the following steps: initialization setting is performed; a target network environment state and historical action feedback are acquired, a meta-cognition evaluation vector is generated, and an action instruction is output; an attack action is executed, an execution result feedback is acquired, if the attack fails, the feedback data, context and failure result are stored in an experience pool, a failure mode is classified and a type label and a solution suggestion are output, a new attack strategy is generated and stored by using a large language model; if the attack succeeds, it is judged whether a test target is achieved, if not, the above process is iterated until the target is achieved or a stop condition is triggered. The application realizes strategy self-evolution through experience learning and strategy evolution, converts failure experience into an executable strategy variant in real time, continuously updates a strategy library as experience accumulates, and enables an intelligent agent to have dual perception of an external environment and an internal state.

Description

Technical Field

[0001] This application relates to the field of cybersecurity technology, and in particular to a metacognitive-driven autonomous attack path detection and self-evolution method and system. Background Technology

[0002] As cyberattack methods become increasingly sophisticated, penetration testing has become a crucial means of proactively discovering system vulnerabilities. Large Language Model (LLM)-based agents, with their powerful reasoning and planning capabilities, offer a new path for automating penetration testing. However, existing LLM-based penetration testing agents suffer from the following technical challenges: 1) Lack of self-evaluation and reflection mechanisms: Most existing intelligent agents are simple feedback models that perceive actions, lacking continuous evaluation and reflection on their internal state (such as knowledge reserves and capability boundaries) and the effectiveness of external strategies. This leads to mechanical path planning, making it difficult to cope with complex and ever-changing network environments, and often failing to make reasonable adjustments when encountering unknown scenarios.

[0003] 2) The evolutionary mechanism is relatively simple. The evolution of existing agents usually relies on offline retraining or pre-set rule updates. When encountering failures in actual penetration testing, it is unable to make flexible code-level strategy adjustments online and autonomously based on real-time failure experience. The lack of efficient online self-evolution capabilities means that the performance of the agent cannot be continuously improved in long-term operation, and manual intervention is required to update the strategy every time a new type of obstacle is encountered. Summary of the Invention

[0004] The purpose of this application is to provide a metacognitive-driven autonomous attack path detection and self-evolution method and system, which can achieve policy self-evolution through experience learning and policy evolution, transform failure experience into executable policy variants in real time, continuously update the policy library with the accumulation of experience, and enable the agent to have dual perception of the external environment and its own internal state.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a metacognitive-driven autonomous attack path detection and self-evolution method, the method comprising: S1. Loading an attack strategy knowledge base and a pre-trained large language model, and receiving preset penetration testing target constraints to complete the penetration testing process initialization; the target constraints include at least: target type, target object identifier, success determination conditions, priority weight, and maximum allowed number of steps; S2. Obtaining the state information of the target network environment and historical execution result feedback; S3. Concatenating the state information, historical execution result feedback, and target constraints into structured prompt words, and converting the structured prompt words into a metacognitive evaluation vector to generate a metacognitive evaluation vector; the metacognitive evaluation vector includes: strategy effectiveness score, knowledge completeness score, resource consumption score, and risk exposure score; the resource consumption score includes: proportion of time used and remaining action budget; the risk exposure score includes: probability of detection. S4. Based on the metacognitive evaluation vector, make a decision on the next attack action and output the action command; S5. Call the security tools to execute the attack action according to the action command and obtain the execution result feedback in real time; S6. Based on the execution result feedback, determine whether the attack action was successfully executed according to the success judgment condition. If not, store the action sequence, environmental context, execution result feedback and failure reason of the attack action as experience into the experience pool, classify and identify the failure mode corresponding to the attack action, output failure type label and solution suggestion, and use this as input to generate a new attack strategy using a pre-trained large language model, update the attack strategy knowledge base, and return to step S1; if yes, proceed to the next step; S7. Use the target constraint condition to determine whether the penetration test goal has been achieved. If yes, the penetration test process ends; if not, repeat steps S2-S6 until the penetration test goal is achieved or the stop condition is triggered.

[0006] Secondly, this application also provides a computer system, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the metacognitive-driven autonomous attack path detection and self-evolution method described in the first aspect.

[0007] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application achieves policy self-evolution through experience learning and policy evolution, transforming failure experiences into executable policy variants in real time. The policy library can be continuously updated as experience accumulates, and the agent has dual perception of the external environment and its own internal state, realizing autonomous decision-making and continuous evolution of the penetration testing agent in complex and unknown environments. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating a metacognitive-driven autonomous attack path detection and self-evolution method provided in this application embodiment. Figure 1 .

[0010] Figure 2 A flowchart illustrating a metacognitive-driven autonomous attack path detection and self-evolution method provided in this application embodiment. Figure 2 .

[0011] Figure 3 This is an internal structure diagram of a computer system provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] Example 1, as Figures 1-2 As shown, this embodiment provides a metacognitive-driven autonomous attack path detection and self-evolution method, the method comprising: S1. Load the attack strategy knowledge base and pre-trained large language model, and receive the preset penetration test target constraints to complete the penetration test process initialization; the target constraints include at least: target type, target object identifier, success determination condition, priority weight and maximum allowed number of steps.

[0015] Furthermore, the target types include: acquiring specified data assets, acquiring specified privilege levels, acquiring the ability to execute commands / implant backdoors on specified hosts, and acquiring specified attack step sequences.

[0016] Optionally, obtaining specified data assets specifically includes: all records in the "users" table of the target database, the administrator password hash, or sensitive files under a specified path (such as / etc / shadow, web.config); obtaining specified privilege levels specifically includes: obtaining root / Administrator privileges, database administrator (DBA) privileges, or web application administrator sessions on the target system; obtaining the ability to execute commands / implant backdoors on a specified host specifically includes: creating a reverse shell connection on the target host, writing specified content to a specified directory, or installing a persistent backdoor program; obtaining a specified attack step sequence specifically includes: completing the complete attack chain of "information gathering → vulnerability exploitation → privilege escalation → data theft" and outputting a penetration test report.

[0017] In practical applications, the initialization steps of the penetration testing process can provide a knowledge base and reasoning ability for subsequent autonomous decision-making.

[0018] S2. Obtain the target network environment status information (environment status snapshot) and historical execution result feedback.

[0019] Furthermore, the status information includes at least one of the following: open port, service fingerprint, and web application path.

[0020] Optionally, historical execution results feedback includes: success, failure, and intermediate status.

[0021] S3. Concatenate the status information, historical execution result feedback, and target constraints into structured prompt words, and convert the structured prompt words into a metacognitive evaluation vector to generate a metacognitive evaluation vector; the metacognitive evaluation vector includes: strategy effectiveness score, knowledge completeness score, resource consumption score, and risk exposure score; the resource consumption score includes: the proportion of time used and the remaining action budget; the risk exposure score includes: the probability of being detected and the number of alarms triggered.

[0022] Optionally, the historical execution result feedback includes historical action sequences and action feedback.

[0023] As an alternative implementation, metacognitive evaluation vectors can be obtained by conversion using a pre-trained large language model.

[0024] Optionally, the strategy effectiveness score includes: current path success rate, expected success probability, and confidence level; the knowledge completeness score includes: known vulnerability coverage, proportion of unknown environments, and knowledge gap identification; the resource consumption score includes: proportion of time used, number of actions used, and remaining action budget; and the risk exposure score includes: probability of detection, number of alerts triggered, and concealment score.

[0025] In practical applications, the metacognitive evaluation vector is generated by a state evaluator based on a large language model. It primarily takes state information (a snapshot of the current environment state) and historical action sequences (windows of length K) from historical execution result feedback as input to generate the metacognitive evaluation vector. The metacognitive evaluation vector contains at least four dimensions: policy effectiveness score, knowledge completeness score, resource consumption score, and risk exposure score, enabling the agent to quantitatively evaluate its own internal state.

[0026] As an optional implementation, the metacognitive evaluation vector also includes: goal progress (used to quantify the distance between the acquired permission level and the target permission level) and team collaboration (used to evaluate collaboration efficiency in multi-agent collaborative scenarios). The core is to maintain the ability to quantify self-evaluation, and the specific dimensions can be flexibly adjusted according to the scenario.

[0027] S4. Based on the metacognitive evaluation vector, make decisions on the next attack action and output action instructions.

[0028] Furthermore, step S4 specifically includes: S41. When the strategy effectiveness score is less than the first threshold or the knowledge completeness score is less than the second threshold, the action instruction corresponding to the next attack action is an exploratory action.

[0029] S42. When the strategy effectiveness score is greater than or equal to the first threshold and the knowledge completeness score is greater than or equal to the second threshold, the action instruction corresponding to the next attack action is an exploitation action.

[0030] S43. When the proportion of time used in the resource consumption score is greater than the third threshold or the remaining action budget is less than the fourth threshold, the action instruction corresponding to the next attack action is the high-confidence shortest path action.

[0031] S44. When the probability of being detected in the risk exposure score is greater than the fifth threshold or the number of alarms triggered is greater than the sixth threshold, the action instruction corresponding to the next attack action will be a low-frequency, low-risk action.

[0032] Further, the first threshold is 0.5; the second threshold is 0.6; the third threshold is 0.8; the fourth threshold is 10; the fifth threshold is 0.5; and the sixth threshold is 3.

[0033] In practical applications, when the strategy effectiveness score is <0.5 or the knowledge completeness score is <0.6, it is judged as "low strategy effectiveness and large knowledge gap," and exploratory actions (such as adding port scans, probing unknown paths, and trying new vulnerability types) are prioritized to expand the scope of information acquisition. When the strategy effectiveness score is ≥0.5 and the knowledge completeness score is ≥0.6, it is judged as "high strategy effectiveness and sufficient knowledge reserves," and the attack continues along the current path, prioritizing exploitative actions (such as exploiting known vulnerabilities and privilege escalation). When the proportion of time used in the resource consumption score is >0.8 or the remaining action budget is <10, a convergence strategy is triggered, prioritizing the shortest path solution with high confidence to avoid resource exhaustion. When the probability of being detected in the risk exposure score is >0.5 or the number of alerts triggered is ≥3, a stealth strategy is triggered, reducing the frequency of actions, prioritizing low-risk, highly stealthy attack actions, and pausing the attack and waiting for a cooldown period if necessary. The entire decision-making process realizes quantitative adaptive strategy planning based on self-awareness, enabling the agent to make quantifiable optimal decisions between exploration and exploitation, and between aggressive and conservative approaches.

[0034] Furthermore, steps S1-S4 are implemented using a metacognitive engine with a single Large Language Model (LLM) through cue word engineering to achieve the four steps of perception, reflection, planning, and action. As an alternative implementation, multiple dedicated LLMs can be used to handle different sub-tasks: for example, an LLM fine-tuned for security testing can be used as the perception unit, and another LLM optimized for reasoning capabilities can be used as the reflection and planning unit, coordinating through message queues or API interfaces. This alternative approach improves the specialized processing capabilities of each stage.

[0035] S5. Invoke security tools to execute attack actions according to the action instructions, and obtain execution result feedback in real time.

[0036] Furthermore, the security tools include at least one of Nmap, SQLmap, and a custom HTTP request library.

[0037] S6. Based on the execution result feedback, determine whether the attack action was successfully executed according to the success criteria. If not, store the action sequence, environmental context, execution result feedback, and failure reason of the attack action as experience in the experience pool, classify and identify the failure mode corresponding to the attack action, output failure type label and solution suggestion, and use this as input to generate a new attack strategy using a pre-trained large language model, update the attack strategy knowledge base, and return to step S1; if yes, proceed to the next step.

[0038] Optionally, new attack strategies are generated by a strategy generator using failure type labels and resolution suggestions as input, based on LLM to create new attack strategies or variations of existing strategies. Generation methods include modifying existing strategy parameters, generating entirely new attack primitive call sequences, or retrieving and combining multiple existing strategies from the strategy library. The generated new strategies are then stored in the strategy library after feasibility scoring for subsequent decision-making.

[0039] Optionally, the experience pool uses a vector database to store experience embeddings, and searches for and matches historical experiences through semantic similarity, thereby improving the efficiency of experience retrieval and reuse.

[0040] As an alternative implementation, a combination of reinforcement learning (RL) and LLM can be used to improve the reliability of policy generation: the policy generated by LLM is used as the initial policy, and reinforcement learning is used to optimize the policy in a simulated environment. The optimized policy parameters are then stored in a policy library. This approach is suitable for scenarios with a large number of simulated training environments.

[0041] Furthermore, the success determination condition is as follows: if any of the following occurs: HTTP response status code indicating rejection, expected response content not matching, tool execution abnormally exiting, or connection interruption, the attack action is determined to have failed; otherwise, the attack action is determined to have succeeded.

[0042] In practical applications, the criteria for determining whether an attack was successfully executed include the following conditions: all of the following conditions must be met for the attack to be considered successful: 1) HTTP response status code indicating rejection: Returning 4xx client errors (such as 403 Forbidden, 404 NotFound) or 5xx server errors (such as 500 Internal Server Error, 503 Service Unavailable) indicates that the attack payload was intercepted or the target service is abnormal.

[0043] 2) Expected response content not matched: The returned HTTP response body does not contain the expected keywords (such as attack success flags, database version information, file list, etc.), meaning that regular expression or semantic matching validation failed.

[0044] 3) Tool execution exits abnormally: The security tool (such as SQLmap, Nmap) returned a non-zero exit code or timed out (the default timeout threshold is 120 seconds), indicating that the attack action was not completed normally.

[0045] 4) Connection interrupted: The target host is unresponsive, the TCP connection is reset (RST), or the firewall actively blocks the connection, and three consecutive retries fail.

[0046] If the failure is determined, experience learning is performed. The sequence of actions, environmental context and execution result of the failure are stored in the experience pool. The experience analyzer based on the Large Language Model (LLM) classifies and identifies the failure mode and outputs failure type labels (such as "WAF interception", "input filtering", "insufficient permissions", "connection blocking") and corresponding solutions (such as "try encoding bypass", "switch User-Agent", "reduce request frequency").

[0047] S7. Use the target constraints to determine whether the penetration test objective has been achieved. If yes, the penetration test process ends; if not, repeat steps S2-S6 until the penetration test objective is achieved or the stop condition is triggered.

[0048] Furthermore, the determination of whether the penetration testing objective has been achieved is based on target constraints, specifically including: 1) For the target type of acquiring a specified data asset, check whether the target data content has been successfully read (e.g., the database query result is not empty and contains the expected fields, and the file content verification is passed). If so, the first sub-target is determined to have been achieved.

[0049] 2) For the target type that obtains a specified permission level, check whether the permission level of the current session has reached the preset target permission level (e.g., by verifying the result returned by the whoami / id command). If so, it is determined that the second sub-target has been achieved.

[0050] 3) For the target type of executing commands / implanting backdoors on a specified host, check whether the command execution echo is consistent with the expected output or whether the backdoor file has been successfully written to the target path and has executable attributes. If so, the third sub-target is determined to have been achieved.

[0051] 4) For obtaining a specified attack step sequence, check whether each attack step in the preset attack sequence has been marked as completed and whether a penetration test report has been generated. If so, the fourth sub-objective is determined to have been achieved.

[0052] 5) When the first, second, third, and fourth sub-objectives are all achieved in order of priority weight, the penetration test objective is considered to have been achieved.

[0053] Furthermore, the stopping condition is reaching the maximum allowed number of steps.

[0054] The technical effects of this application are as follows: This application presents an agent self-evolution method based on a large language model and metacognitive architecture for autonomous network attack path detection and planning. It achieves policy self-evolution through experience learning and policy evolution, transforming failure experiences into executable policy variants in real time. The policy library can be continuously updated with the accumulation of experience, and the agent has dual perception of the external environment and its own internal state, realizing autonomous decision-making and continuous evolution of the penetration testing agent in complex and unknown environments.

[0055] Example 2: This example provides a computer system, which can be a server or a terminal, and its internal structure diagram can be as follows. Figure 3 As shown, the computer system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements the methods described above.

[0056] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer system to which the present application is applied. A specific computer system may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0057] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0058] The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited thereto.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0060] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A metacognitive-driven autonomous attack path detection and self-evolution method, characterized in that, The method includes: S1. Load the attack strategy knowledge base and pre-trained large language model, and receive the preset penetration test target constraints to complete the penetration test process initialization; the target constraints include at least: target type, target object identifier, success determination condition, priority weight, and maximum allowed number of steps; S2. Obtain the status information of the target network environment and historical execution result feedback; S3. Concatenate the status information, historical execution result feedback, and target constraints into structured prompt words, and convert the structured prompt words into a metacognitive evaluation vector; the metacognitive evaluation vector includes: strategy effectiveness score, knowledge completeness score, resource consumption score, and risk exposure score; the resource consumption score includes: the proportion of time used and the remaining action budget; the risk exposure score includes: the probability of being detected and the number of alarms triggered; S4. Based on the metacognitive evaluation vector, make a decision on the next attack action and output the action command; S5. Invoke security tools to execute attack actions according to the action instructions, and obtain execution result feedback in real time; S6. Based on the execution result feedback, determine whether the attack action was successfully executed according to the success criteria. If not, store the action sequence, environmental context, execution result feedback, and failure reason of the attack action as experience in the experience pool, classify and identify the failure mode corresponding to the attack action, output failure type label and solution suggestion, and use this as input to generate a new attack strategy using a pre-trained large language model, update the attack strategy knowledge base, and return to step S1; if yes, proceed to the next step. S7. Use the target constraints to determine whether the penetration test objective has been achieved. If yes, the penetration test process ends; if not, repeat steps S2-S6 until the penetration test objective is achieved or the stop condition is triggered.

2. The method for autonomous attack path detection and self-evolution based on metacognitive drive according to claim 1, characterized in that, Based on metacognitive evaluation vectors, decisions are made regarding the next attack action, and action instructions are output, specifically including: When the strategy effectiveness score is less than the first threshold or the knowledge completeness score is less than the second threshold, the action instruction corresponding to the next attack action will be an exploratory action. When the strategy effectiveness score is greater than or equal to the first threshold and the knowledge completeness score is greater than or equal to the second threshold, the action instruction corresponding to the next attack action is an exploitation action. When the proportion of time used in the resource consumption score is greater than the third threshold or the remaining action budget is less than the fourth threshold, the action instruction corresponding to the next attack action is the high-confidence shortest path action. When the probability of being detected in the risk exposure score is greater than the fifth threshold or the number of alerts triggered is greater than the sixth threshold, the action instruction corresponding to the next attack action will be a low-frequency, low-risk action.

3. The metacognitive-driven autonomous attack path detection and self-evolution method according to claim 2, characterized in that, The first threshold is 0.5; the second threshold is 0.6; the third threshold is 0.8; the fourth threshold is 10; the fifth threshold is 0.5; and the sixth threshold is 3.

4. The method for autonomous attack path detection and self-evolution based on metacognitive drive according to claim 1, characterized in that, The success criteria are as follows: if any of the following occurs: HTTP response status code indicating rejection, expected response content not matching, tool execution abnormally exiting, or connection interruption, the attack action is deemed to have failed; otherwise, the attack action is deemed to have succeeded.

5. The method for autonomous attack path detection and self-evolution based on metacognitive drive according to claim 1, characterized in that, The target types include: acquiring specified data assets, acquiring specified privilege levels, acquiring commands / implanting backdoors on specified hosts, and acquiring specified attack step sequences.

6. The metacognitive-driven autonomous attack path detection and self-evolution method according to claim 5, characterized in that, Determining whether the penetration testing objective has been achieved using target constraints includes: For the target type of acquiring a specified data asset, check whether the target data content has been successfully read. If so, the first sub-target is considered to have been achieved. For the target type that obtains a specified permission level, check whether the permission level of the current session has reached the preset target permission level. If so, it is determined that the second sub-goal has been achieved. For obtaining the target type of executing commands / implanting backdoors on a specified host, check whether the command execution echo is consistent with the expected output or whether the backdoor file has been successfully written to the target path and has executable attributes. If so, it is determined that the third sub-target has been achieved. If a specified attack step sequence is obtained, check whether each attack step in the preset attack sequence has been marked as completed and whether a penetration test report has been generated. If so, the fourth sub-objective is determined to have been achieved. When the first, second, third, and fourth sub-objectives are all achieved in order of priority weight, the penetration test objective is considered to have been achieved.

7. The method for autonomous attack path detection and self-evolution based on metacognitive drive according to claim 1, characterized in that, The status information includes at least one of the following: open port, service fingerprint, and web application path.

8. The method for autonomous attack path detection and self-evolution based on metacognitive drive according to claim 1, characterized in that, The security tools include at least one of Nmap, SQLmap, and a custom HTTP request library.

9. The metacognitive-driven autonomous attack path detection and self-evolution method according to claim 1, characterized in that, The stopping condition is reaching the maximum allowed number of steps.

10. A computer system, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the metacognitive-driven autonomous attack path detection and self-evolution method according to any one of claims 1-9.

Images