Agent-based security verification method, device, equipment and medium

Abstract

Embodiments of the application disclose a security verification method and device based on an agent, equipment and a medium. The method constructs a multi-agent collaborative architecture of a security classification agent, a business processing agent and an output review agent, and performs a full-process step of cascade detection, security check, business processing, output review and audit log recording, thereby performing all-round and multi-level security protection on a financial case verification process from input to output, improving attack interception rate, reducing false positive rate, and meeting real-time, compliance and traceability requirements of a financial scenario.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of artificial intelligence technology, and in particular to a security verification method, apparatus, device, and medium based on intelligent agents. Background Technology

[0002] In the field of financial risk control, with the rapid development of artificial intelligence technology, intelligent agents have been widely applied in core business scenarios such as intelligent customer service, credit approval, and anti-fraud investigation. These AI systems, through deep integration with large language models, can efficiently handle complex tasks such as financial case inquiries, risk assessments, and report generation, significantly improving business processing efficiency. However, the financial industry, as a special sector with strong regulatory and high security requirements, has business systems that directly serve a massive number of users and store large amounts of sensitive customer information, transaction data, and core risk control rules.

[0003] Therefore, extremely high requirements are placed on the security protection of AI systems. How to effectively prevent malicious attacks against intelligent agents, such as prompt word injection, rule detection, and theft of sensitive information, has become a key technical problem that financial institutions urgently need to solve. Summary of the Invention

[0004] In view of this, embodiments of this application provide at least one security verification method, apparatus, device, and medium based on intelligent agents.

[0005] The technical solution of this application embodiment is implemented as follows: On one hand, embodiments of this application provide a security verification method based on intelligent agents, the method comprising: The received user query is input into the security classification agent, and the security of the user query is detected by the cascaded detection mechanism of the security classification agent to obtain a security judgment result. Based on the security determination result, if the user query is an insecure query, a preset interception response is executed and a security event log is recorded; if the user query is a secure query, the user query is forwarded to the business processing intelligent agent. The business processing intelligent agent executes the financial case verification logic to obtain a business response; The business response is input to the output review agent, which performs sensitivity and compliance checks on the business response to obtain the output review result. Based on the output review result, if the output review result is passed, the business response is output; if the output review result is not passed, the business response is replaced with a preset security response and then output, and all security events are recorded in the audit log module.

[0006] On the other hand, embodiments of this application provide a security verification device based on an intelligent agent, the method comprising: The processing module is used to input the received user query into the security classification agent, and to perform security detection on the user query through the cascading detection mechanism of the security classification agent to obtain a security judgment result; Based on the security determination result, if the user query is an insecure query, a preset interception response is executed and a security event log is recorded; if the user query is a secure query, the user query is forwarded to the business processing intelligent agent. The business processing intelligent agent executes the financial case verification logic to obtain a business response; The output module is used to input the business response to the output review agent, and the output review agent performs sensitivity checks and compliance checks on the business response to obtain the output review result. Based on the output review result, if the output review result is passed, the business response is output; if the output review result is not passed, the business response is replaced with a preset security response and then output, and all security events are recorded in the audit log module.

[0007] In another aspect, embodiments of this application provide a computer device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-described agent-based security verification method.

[0008] In another aspect, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described agent-based security verification method.

[0009] In another aspect, embodiments of this application provide a computer program including computer-readable code. When the computer-readable code is run in a computer device, the processor in the computer device executes some or all of the steps in the above-described agent-based security verification method.

[0010] In another aspect, embodiments of this application provide a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described agent-based security verification method.

[0011] This application embodiment constructs a multi-agent collaborative architecture consisting of a security classification agent, a business processing agent, and an output review agent, and executes the entire process of cascading detection, security checks, business processing, output review, and audit log recording. From input to output, it provides comprehensive and multi-layered security protection for the financial case verification process, improving the attack interception rate, reducing the false alarm rate, and meeting the real-time, compliance, and traceability requirements of financial scenarios.

[0012] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this application. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0014] Figure 1 A schematic diagram illustrating the implementation process of a security verification method based on an intelligent agent, provided in an embodiment of this application; Figure 2 This application provides a schematic diagram of the composition structure of a security verification device based on an intelligent agent. Figure 3 This is a schematic diagram of the hardware entity of a computer device provided in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0017] The terms “first / second / third” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first / second / third” may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of this application.

[0019] This application provides a security verification method based on an intelligent agent, which can be executed by the processor of a computer device. The computer device can refer to a server, laptop computer, tablet computer, desktop computer, smart TV, set-top box, mobile device (such as a mobile phone, portable video player, personal digital assistant, dedicated messaging device, portable gaming device), or other similar computer equipment. Figure 1 This application provides a schematic diagram of the implementation process of a security verification method based on an intelligent agent, as illustrated in the embodiments below. Figure 1 As shown, the method includes: Step 101: Input the received user query into the security classification agent, and perform security detection on the user query through the cascading detection mechanism of the security classification agent to obtain a security determination result.

[0020] In this embodiment, a user query is information submitted by a user to inquire about or request an operation, such as a piece of text or a question. The security classification agent is a component specifically responsible for security detection; its core function is to identify and block user queries containing malicious intent. The cascaded detection mechanism is a multi-level, progressive detection process employed within the security classification agent, consisting of multiple detectors with varying detection accuracy and speed connected in series. Security detection is the process of analyzing user queries to determine whether they contain malicious intent or sensitive information. The security determination result is the final conclusion, after completing the security detection, regarding whether the user query is unsafe or safe. Confidence level is a quantitative value representing the reliability of the security determination result; a higher value indicates a more credible determination result.

[0021] The system receives a query from the user. This query is then input into a dedicated security classification agent responsible for security detection. This agent initiates its built-in cascading detection mechanism, which sequentially performs security checks on the user query through multiple detection levels. This process begins with the fastest rule matching, proceeds to semantic analysis, and may eventually involve using a large language model for deep understanding. After completing all levels of detection, the system synthesizes the results from each level to generate a final security judgment, such as secure or insecure, and also outputs a confidence score representing the reliability of this judgment.

[0022] Step 102: Based on the security determination result, if the user query is an insecure query, execute the preset interception response and record the security event log; if the user query is a secure query, forward the user query to the business processing intelligent agent.

[0023] In this embodiment, the security determination result is the final conclusion regarding whether a user query is secure or insecure. An insecure query is a user query that, after security testing, is determined to contain malicious intent, offensive, or sensitive information. An interception response is a pre-defined standard reply to a query deemed insecure, used to deny service and inform the user. A security event log is a file or data record that documents detailed information about all security-related events. A secure query is a user query that, after security testing, is determined to be harmless and its business request can continue to be processed. The business processing intelligent agent is another component specifically responsible for executing core financial business logic.

[0024] Based on the security assessment result, branching processing occurs. If the assessment result indicates that the user query is insecure, a pre-set interception response will be executed immediately, such as returning a standard message like "Sorry, we cannot process your request," and an automatic record containing detailed information such as the user ID, query content, and threat type will be recorded in the security event log. Conversely, if the assessment result indicates that the user query is secure, the query will be forwarded to the next business processing agent responsible for handling core business logic, so that subsequent business logic processing can continue.

[0025] Step 103: The business processing intelligent agent executes the financial case verification logic to obtain a business response.

[0026] In this embodiment, the business processing agent is a component specifically responsible for executing the core business logic of financial case verification. The financial case verification logic comprises a series of business rules and calculation processes for handling financial cases. Business response is the result information generated and output by the business processing agent after completing all business logic. Case query is the process by which the business processing agent retrieves relevant case information from the database based on information queried by the user. Risk control rule matching is the process by which the business processing agent compares case information with preset risk control rules to identify potential risks. Risk analysis report generation is the process by which the business processing agent automatically organizes and forms a document containing risk scores, analysis conclusions, etc., after completing case query and risk control rule matching.

[0027] Upon receiving and forwarding a security query, the business processing agent immediately begins executing the financial case verification logic, which includes several sub-steps. First, a case query is performed, retrieving corresponding case information, user data, and transaction records from the database based on the user or case identifier in the query. Next, risk control rule matching is performed, comparing the retrieved case information with pre-configured risk control rules, such as blacklist matching and transaction amount thresholds. Based on the results, a risk score is calculated using rules and models, and a risk analysis report containing risk analysis and handling suggestions is automatically generated. The output of the entire process is the business response.

[0028] Step 104: Input the business response into the output review agent, and perform sensitivity and compliance checks on the business response through the output review agent to obtain the output review result.

[0029] In this embodiment, the business response is the result information generated by the business processing agent after completing the business logic. The output review agent is a component specifically responsible for performing a final security check on the business response. Sensitivity checking is the process of analyzing the content of the business response to determine whether it contains sensitive internal information. Compliance checking is the process of analyzing the content of the business response to determine whether it complies with financial regulatory provisions and compliance requirements. The output review result is the final conclusion on whether the business response passes or fails after completing the sensitivity and compliance checks.

[0030] The business response is input to the output review agent. Upon receiving the response, the agent performs two checks. The first is a sensitivity check, scanning the text to detect whether it contains sensitive information that should not be disclosed, such as prompts, source code, API keys, or risk control thresholds. The second is a compliance check, examining whether the response contains any suggestions, promises, or data disclosures that do not comply with financial regulatory requirements. After completing these two checks, a comprehensive evaluation is conducted to arrive at a final output review result, such as pass or fail.

[0031] Step 105: Based on the output review result, if the output review result is passed, output the business response; if the output review result is not passed, replace the business response with a preset security response and output it, and record all security events to the audit log module.

[0032] In this embodiment, the output review result is the final conclusion regarding whether the business response passes or fails. The security response is a pre-defined standard reply content used to replace business responses that fail the output review. The audit log module is a component specifically designed for centralized storage and management of all security event records.

[0033] Final processing is performed based on the output review results. If the output review passes, the original business response is directly output to the user as the final result. If the output review fails, the original business response is not output; instead, it is replaced with a preset, secure standard response that does not disclose any sensitive information, such as "Sorry, this information cannot be provided," and this security response is output. Regardless of the review result, all security events related to this processing, including user ID, query content, judgment result, threat type, timestamp, etc., are recorded in a centralized audit log module for subsequent auditing and analysis.

[0034] This application embodiment constructs a multi-agent collaborative architecture consisting of a security classification agent, a business processing agent, and an output review agent, and executes the entire process of cascading detection, security checks, business processing, output review, and audit log recording. From input to output, it provides comprehensive and multi-layered security protection for the financial case verification process, improving the attack interception rate, reducing the false alarm rate, and meeting the real-time, compliance, and traceability requirements of financial scenarios.

[0035] Optionally, step 101 includes: Step 1011: Perform keyword matching detection on the user query through a rule filter. The rule filter contains a preset number of financial-specific malicious patterns. If the match is successful, output the first security judgment result as unsafe and the first confidence level.

[0036] In this embodiment, the rule filter is a built-in program module used for rapid keyword matching and detection of user-input text. Financial-specific malicious patterns are a predefined series of specific strings or rules specifically designed to target common attack methods in financial scenarios, such as your prompt words, backend code, risk control rules, and bypass techniques.

[0037] A rule filter is activated. This filter receives the user query as input and immediately compares it with a preset number of financial-specific malicious patterns stored internally. If any part of the user query matches any of these patterns, the query is determined to be insecure, and a first security judgment result of "insecure" is generated, along with a preset first confidence level assigned to this judgment result.

[0038] Step 1012: If the first confidence score of the rule filter is lower than the preset first confidence score threshold, then the small model classifier is invoked to perform semantic analysis on the user query. The small model classifier is a Chinese pre-trained language model that has been supervised and fine-tuned by a preset amount of financial security labeled corpus, and outputs a second security judgment result and a second confidence score.

[0039] In this embodiment, the small model classifier is a pre-trained Chinese language model with 100M to 400M parameters, such as BERT-base-chinese or RoBERTa-base. The financial security-annotated corpus is manually annotated text data related to financial security, specifically used to train the model. The preset quantity in this scheme refers to at least 1000 entries. Semantic analysis refers to the model's ability to understand the deeper meaning and intent of the text, rather than simply matching keywords.

[0040] When the initial confidence level of the rule filter falls below a preset initial confidence threshold, it indicates that the keyword matching result is uncertain and may be susceptible to complex attacks such as semantic spoofing. In this case, a small model classifier is automatically invoked to perform a deeper semantic analysis of the user query. Since this small model classifier has undergone supervised fine-tuning with at least 1000 financial security-annotated corpora, it can identify malicious intent disguised through synonym substitution or sentence structure transformation. After analysis, a second security determination result and the corresponding second confidence level are output; the inference time for this step is less than 50 milliseconds.

[0041] Step 1013: If the second confidence score of the small model classifier is lower than the preset first confidence score threshold, the large language model classifier is invoked to perform deep semantic analysis on the user query. The large language model classifier identifies multiple threat types through preset prompt words containing security classification instructions and outputs a third security judgment result, threat type and third confidence score.

[0042] In this embodiment, the large language model classifier is a language model with a much larger number of parameters than the small model, possessing stronger semantic understanding and reasoning capabilities. Preset prompt words are fixed instruction texts designed to guide the large language model in performing specific security classification tasks. Threat types refer to the specific attack categories that can be identified, such as prompt word injection attacks, code extraction attacks, rule probing attacks, etc.

[0043] If the second confidence score of the small model classifier is still lower than the preset first confidence score threshold, it indicates that the semantics of the query are very complex or hidden, and the small model cannot make a high-confidence judgment. In this case, the large language model classifier will be invoked for the most in-depth semantic analysis. The user query, along with a preset prompt containing security classification instructions, will be input into the large language model to guide it in identifying multiple threat types. After analysis, the large language model will output a third security judgment result, the specific threat type, and a broader third confidence score.

[0044] Step 1014: Combining the first security determination result, the second security determination result, and the third security determination result, if the security determination result at any level is unsafe, the target security determination result is unsafe, and the highest value among the first confidence level, the second confidence level, and the third confidence level is taken as the confidence level.

[0045] In this embodiment, comprehensive judgment refers to the process of summarizing and logically operating on the judgment results from the rule filter, small model classifier, and large language model classifier. The target security judgment result is the final output, used to determine whether to block the query. Confidence level is a quantitative score of the reliability of the final judgment result.

[0046] The system receives three security determination results from steps 1011, 1012, and 1013, and synthesizes them using a veto logic: if any level of the rule filter, small model classifier, or large language model classifier determines the target to be unsafe, the system determines the target to be unsafe. Simultaneously, the highest confidence score among the three is selected as the final output confidence score for subsequent processes. This comprehensive determination method ensures that security interception is triggered at any level of detection, achieving maximum protection coverage.

[0047] This application embodiment improves the attack interception rate by employing a three-level cascaded detection mechanism consisting of a rule filter, a small model classifier, and a large language model classifier, and by comprehensively judging the results at each level.

[0048] Optionally, step 102 includes: Step 1021: Generate a preset standard rejection message as the interception response, and return the standard rejection message to the user terminal.

[0049] In this embodiment of the application, the standard rejection message is a pre-set fixed text that is used to send a uniform rejection reply to the user when an insecure query is detected, without containing any specific business information.

[0050] When a user query is deemed unsafe, a predefined text template is invoked. This template contains a standard rejection message such as "Hello, your query cannot be processed. Please re-enter," or similar message. This standard rejection message is generated and used as the official response to the user's query. Subsequently, this standard rejection message is sent back to the user's terminal that initiated the query via a network interface, clearly informing the user that their request has been rejected without revealing any details about the reason for the rejection or the internal mechanisms.

[0051] Step 1022: Obtain the threat type and confidence level output by the security classification agent, and record the security event log to the audit log module. The security event log includes the user identifier, the user query, the threat type, the confidence level, and the current timestamp.

[0052] In this embodiment, the security event log is a detailed record file or database entry used to record all security-related events. The user identifier is a unique identification code assigned to each user. The user query is the raw text content submitted by the user. The threat type is a specific classification of the detected attack behavior, such as a prompt word injection attack. The confidence level is a numerical measure of the credibility of its security judgment result, typically ranging from 0 to 1. The current timestamp records the precise date and time information when the event occurred.

[0053] While executing step 1021, the threat type and corresponding confidence level identified in this security detection are extracted from the security classification agent performing the security detection. This information is then combined with the user's unique identifier, the original query text submitted by the user, and the current timestamp when the event was recorded to form a complete security event log entry. The write interface of the audit log module is then called to persistently store this entry in a specified log storage location, such as a database or log file, for subsequent auditing and analysis.

[0054] Step 1023: Terminate subsequent processes, prohibit forwarding the user query to the business processing intelligent agent, and prohibit the execution of the financial case verification logic and the inspection of the output review intelligent agent.

[0055] In this embodiment, the business processing agent is a dedicated program module responsible for executing the core verification logic of financial cases. The financial case verification logic comprises a series of business rules and calculation steps designed to handle specific financial cases. The output review agent is a dedicated program module responsible for performing a secondary security check on the business processing results.

[0056] After logging the security incident, the processing flow of the current user's query is immediately interrupted. An internal status flag is set to explicitly prohibit the sending of the user's query data packet to the business processing agent, thereby preventing it from executing any business logic related to financial cases, such as queries, risk control rule matching, or risk report generation. Simultaneously, subsequent calls to the output review agent are also prohibited; that is, no further secondary sensitivity checks are needed on any output content because the business process itself has been terminated. Through this series of operations, it is ensured that any query deemed unsafe cannot access core business data or subsequent processing steps, achieving strict isolation for security protection.

[0057] This application embodiment achieves immediate blocking of insecure queries, full-link traceability, and complete isolation of business processes by generating and returning standard rejection statements, recording detailed security event logs, and forcibly terminating all subsequent business processes, thereby enabling closed-loop management and zero-tolerance execution of security protection.

[0058] Optionally, step 103 includes: Step 1031: Analyze the intent of the user query and extract business parameters from the user query. The business parameters include case identifier and user information.

[0059] In this embodiment, a user query is text data input by a user for querying or manipulating information about financial cases; intent is the specific operational purpose that the user wants to perform, identified by analyzing the text content of the user query; business parameters are specific data items extracted from the user query for executing business logic; case identifier is a number or code used to uniquely identify a financial case; and user information is identity data related to the user who initiated the query.

[0060] Receive user query text verified as secure by the security classification agent. Perform natural language processing on the text to parse out the user's desired action, such as querying case details or conducting a risk assessment. Extract the business parameters required for execution from the user query. These parameters specifically include a case identifier for locating a specific financial case and user information for identifying the query initiator.

[0061] Step 1032: Query the corresponding case information and user transaction records according to the business parameters, match the case information with the preset risk control rules, and obtain the risk control rule matching result.

[0062] In this embodiment of the application, case information is detailed data stored in a file corresponding to a case identifier, including the basic information of the case, processing records, etc.; user transaction records are the user's past financial transaction history data; risk control rules are preset logical conditions or thresholds used to determine whether a transaction or case has risks; risk control rule matching results are conclusions generated by comparing case information with risk control rules, indicating whether the rules are triggered.

[0063] Using the extracted business parameters, namely case identifier and user information, the system queries the database for corresponding case information and the user's transaction records. This retrieved case information is then compared one by one with pre-configured risk control rules to determine whether the case characteristics match any rule's defined abnormal or high-risk pattern. Finally, a risk control rule matching result is generated, which may include a list of matched rules and the degree of matching.

[0064] Step 1033: Calculate the risk score based on the risk control rule matching result and the preset risk scoring model, wherein the risk scoring model combines a rule engine and a machine learning model.

[0065] In the embodiments of this application, the risk score is a numerical value calculated to quantify the risk level corresponding to a case or query; the risk score model is an algorithm or function used to calculate the risk score, which combines two technologies: rule engine and machine learning model.

[0066] The system receives the generated risk control rule matching results and invokes a preset risk scoring model. This model is not a single algorithm, but a composite model that integrates both rule engine and machine learning model methods. The rule engine scores based on the triggered risk control rules and their weights, while the machine learning model assesses the risk of cases based on patterns trained on historical data. The results of these two calculations are integrated to finally calculate a comprehensive risk score.

[0067] Step 1034: Generate a risk analysis report based on the risk score and the case information, and output the risk analysis report as the business response.

[0068] In this embodiment, the risk analysis report is a structured document generated based on risk scores and case information, containing risk assessment conclusions and detailed evidence; the business response is the final result prepared to be output to the user after processing the user query.

[0069] Based on the risk score and case information, a structured risk analysis report is automatically generated. This report includes basic case information, risk score, triggered main risk control rules, and risk analysis conclusions. This risk analysis report serves as the final business response to this business request and is prepared to be passed to the output review agent for subsequent sensitivity checks.

[0070] This application embodiment achieves efficient, automated, and multi-dimensional risk assessment of financial cases by parsing user queries into business parameters, querying case information and transaction records, applying risk control rules and a composite risk scoring model for calculation, and generating a structured risk analysis report. This improves the accuracy of business processing and the efficiency of risk identification.

[0071] Optionally, step 104 includes: Step 1041: Perform a sensitivity check on the business response. The sensitivity check detects whether the business response contains sensitive information through rule matching and semantic analysis. The sensitive information includes prompt words, source code, application programming interface keys, and risk control rule thresholds.

[0072] In this embodiment, the business response is the result data generated after the business processing is completed. Sensitivity checking is a security detection operation performed on data content to identify whether it contains specific types of sensitive information. Rule matching is a detection method that compares data content using a preset keyword or pattern library. Semantic analysis is a detection method that identifies potential risks by understanding the deeper meaning of text. Sensitive information refers to data that requires strict protection and whose leakage may cause security risks or violate regulations. Prompt words are preset instructions used internally to guide the behavior or output of artificial intelligence models. Source code is a set of program instructions running at the underlying level. The application programming interface key is a credential used to verify the identity of the service caller. Risk control rule thresholds are numerical limits used to determine the risk level.

[0073] Receives a business response generated by the business processing module. Initiates a sensitivity check process, which integrates two detection technologies: firstly, rule matching technology compares the business response content with a predefined sensitive information keyword database; secondly, semantic analysis technology understands the deeper meaning of the business response text to identify expressions that are disguised by words but actually point to sensitive information. The core objective of this check is to detect whether the business response contains any sensitive information, which specifically includes, but is not limited to, prompts, program source code, application programming interface keys used to verify service identity, and risk control rule thresholds.

[0074] Step 1042: Perform a compliance check on the business response. The compliance check examines whether the business response complies with financial regulatory requirements, whether it contains inappropriate advice or promises, and whether the data disclosure is compliant.

[0075] In this application embodiment, compliance checking is a compliance verification operation performed on data content to assess whether it complies with external or internal regulatory requirements. Financial regulatory requirements are laws, regulations, and industry standards formulated by financial industry regulatory agencies that must be followed. Inappropriate advice or promises are statements that may mislead users or expose them to legal risks. Data disclosure is information displayed externally in the response.

[0076] After completing the sensitivity check, a compliance check is performed on the same business response. Using a pre-defined rule base and models, each business response is verified to ensure it meets the regulatory requirements of the financial industry, such as checking for any legally prohibited statements. Simultaneously, the response is examined for any statements that could constitute inappropriate advice or promises, potentially leading users to make incorrect decisions or subjecting the operator to legal liability. Furthermore, the data disclosed in the response is reviewed to ensure it complies with relevant regulations on data use and privacy protection.

[0077] Step 1043: If both the sensitivity check and the compliance check pass, the output review result is passed.

[0078] In this embodiment, after completing the sensitivity check and compliance check separately, a logical judgment is made by combining the results of the two checks. If it is determined that the business response does not contain any sensitive information in the sensitivity check and does not contain any illegal content in the compliance check, meaning both checks are passed, then the result of this output review is deemed passed. This means that the business response is considered safe and compliant and can be normally output to the user.

[0079] Step 1044: If either the sensitivity check or the compliance check fails, the output review result is "failed".

[0080] In this embodiment, contrary to the logic of step 1043, during the comprehensive judgment, if either the sensitivity check or the compliance check fails—that is, the business response is detected to contain sensitive information or violate compliance requirements, or both fail—then the result of this output review is determined to be unsuccessful. This determination means that the business response has security risks or compliance issues, and its direct output will be prevented, triggering a preset subsequent processing flow, such as replacing the response content with a secure default reply.

[0081] This application embodiment employs a dual verification mechanism that performs sensitivity and compliance checks on business responses, and determines whether the output review passes based on the verification results. This enables full-process security control of the output content, effectively prevents the leakage of sensitive information, and ensures that business responses comply with financial regulatory requirements.

[0082] Optionally, step 105 includes: Step 1051: Record the security event logs corresponding to all intercepted user queries. The security event logs include user identifier, query content, threat type, confidence level, and timestamp.

[0083] In this embodiment, a user query is a request text submitted by a user to obtain information or perform an operation. A security event log is a data set used to record events related to security risks. A user identifier is an identity code used to uniquely distinguish different users. The query content is the original request text data submitted by the user. The threat type is the specific category of security attack identified, such as prompt word injection attacks, code extraction attacks, or rule probing attacks. Confidence level is a quantified probability value given for the accuracy of the judgment result. A timestamp is a data marker recording the precise moment the event occurred.

[0084] After a security check is performed, a logging process is automatically triggered for all user queries deemed insecure and for which an intercept response was initiated. The security event logs corresponding to these intercepted user queries are recorded in a structured manner, including user identifiers to distinguish different users, the original query content submitted by the user, the specific threat type identified, the confidence level of the security assessment result, and the precise timestamp of the event. This method constructs a complete audit log for all intercepted attacks.

[0085] Step 1052: Record the pass logs for all user queries that pass the security test. The pass logs include user identifier, query content, business response summary and timestamp.

[0086] In this embodiment, the log is a data set used to record all user queries that successfully pass the security check. The business response summary is a concise description of the business result generated after processing the user query.

[0087] For all user queries that pass security checks and enter the business processing flow, another set of logging logic is executed after the business processing is completed. This log records the corresponding user queries, including a user identifier, the original query content submitted by the user, a summary of the generated business response, and a timestamp of the event. This method is used to construct the processing records for all normal business requests.

[0088] Step 1053: Record the output review event log, which includes the output review result, the sensitive information type, and the replaced response content.

[0089] In this embodiment, the output review event log is a dataset used to record the results of sensitivity checks on business response content. The output review result is the final conclusion given after reviewing the business response content, such as pass or block. Sensitive information type is the specific category of information that has been identified as being leaked, such as prompt words, source code, or risk control rule thresholds. The replaced response content is the security response text used to replace the original response when the original response fails the review.

[0090] After the output review process is executed, regardless of whether the review result is approval or blocking, a corresponding output review event log is generated and recorded. This log details the output review result, the types of sensitive information discovered during the review process, and the response text used to replace the original content when the response is blocked. This method constructs a complete monitoring record for output security.

[0091] Step 1054: Store the security event log, the pass log, and the output review event log in a traceable database, supporting queries by user ID, time range, and threat type.

[0092] In this embodiment, the traceable database is one that possesses data integrity, immutability, and historical query capabilities. The threat type refers to the specific category of security attack identified, such as prompt word injection attacks, code extraction attacks, or rule probing attacks.

[0093] All generated log data, including security event logs, pass logs, and output audit event logs, are stored uniformly in a highly secure and traceable database. This database supports efficient queries by multiple dimensions such as user ID, time range, and threat type, enabling security auditors or administrators to quickly locate, analyze, and trace any historical security events or business requests.

[0094] This application embodiment constructs a comprehensive audit log system that includes security event logs, pass logs, and output review event logs, and stores it in a traceable database. This achieves complete recording and traceable querying of the entire process of security verification for financial cases, resulting in traceable security events, transparent auditing processes, and continuous optimization of the security model.

[0095] Optionally, the method further includes: Step 201: Record the security event logs corresponding to all intercepted user queries. The security event logs include user identifier, query content, threat type, confidence level, and timestamp.

[0096] In this embodiment, the security event log is a log file that records information related to all user queries deemed unsafe. The user identifier is an identity code used to uniquely identify a user. The query content is the original question or instruction text submitted by the user. The threat type is the specific attack category identified in the user query, such as a prompt word injection attack or a code extraction attack. The confidence level is the degree of certainty regarding the security determination result, typically represented by a value between 0 and 1. The timestamp is a marker recording the specific time the event occurred.

[0097] The system performs security checks on user-submitted queries. When a query is deemed insecure, a logging mechanism is triggered. This creates a security event log entry and extracts key information related to the user's query. The user's identity code is recorded as the user identifier, the original text submitted by the user is recorded as the query content, the detected attack category is recorded as the threat type, the level of certainty regarding the determination is recorded as the confidence level, and the precise time of the event is recorded as a timestamp. All this information is combined into a complete log entry.

[0098] Step 202: Record the pass logs for all user queries that pass the security test. The pass logs include user identifier, query content, business response summary and timestamp.

[0099] In this embodiment, the log is a log file that records all user queries deemed safe. The user identifier is an identity code used to uniquely identify a user. The query content is the original question or instruction text submitted by the user. The business response summary is a key information extraction from the final result returned to the user after business processing is completed. The timestamp is a marker recording the specific time the event occurred.

[0100] After performing a security check on a user-submitted query, if the result is deemed secure, another logging mechanism is triggered. This creates a pass log entry. The user's identity code is recorded as the user identifier, and the original text submitted by the user is recorded as the query content. The system waits for the business processing flow to complete and extracts key information from the final generated business response to form a business response summary, such as a risk score or case status. This summary, along with the precise timestamp of the event, is recorded in the pass log entry.

[0101] Step 203: Record the output review event log, which includes the output review results, sensitive information types, and the replaced response content.

[0102] In this embodiment, the output review event log is a log file that records all information related to the output review process of the business response. The output review result is the final conclusion after sensitivity checking of the business response, such as pass or block. Sensitive information type is the specific sensitive data category identified from the blocked business response, such as prompt words or source code. The replaced response content is the standardized security prompt text used to replace and return to the user after the original response is blocked.

[0103] After business processing is completed, the generated business response undergoes output review. Regardless of whether the review result is approval or blocking, an output review event log is created. If the review result is blocking, the specific sensitive information type that caused the blocking is identified, such as warning words or risk control rule thresholds. Simultaneously, the security warning text used to replace the original response is recorded as the replaced response content. All this information, including the output review result, sensitive information type, and the replaced response content, is recorded in this log.

[0104] Step 204: Store the security event log, the pass log, and the output review event log in a traceable database, supporting queries by user ID, time range, and threat type.

[0105] In this embodiment, the security event log is a log file that records all intercepted user query information. The pass log is a log file that records all user query information deemed safe. The output review event log is a log file that records all information related to the output review process of business responses. The traceable database is persistent storage for storing logs; its design ensures that once data is written, it can be audited and traced, and cannot be arbitrarily modified or deleted. The user identifier is an identity code used to uniquely identify a user. The time range is a start and end time period used to filter log records. The threat type is the specific attack category identified from the user query.

[0106] Security event logs, pass logs, and output review event logs are all written to a unified, traceable database. This database design ensures the integrity and immutability of the log data. A query interface is then provided, allowing security analysts or administrators to retrieve and view relevant historical log records from this database by specifying user IDs, time ranges, or threat types, for security event analysis and optimization.

[0107] This application embodiment achieves end-to-end traceability and queryability of all security-related information by uniformly recording three types of logs: security events, events passed through, and events output for review, and storing them in a traceable database. This provides a complete data foundation for security auditing, attack pattern analysis, and continuous model optimization.

[0108] Based on the foregoing embodiments, this application provides a security verification device based on an intelligent agent. The device includes various units and modules included in each unit, which can be implemented by a processor in a computer device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0109] Figure 2 This application provides a schematic diagram of the composition structure of a security verification device based on an intelligent agent, as shown in the embodiments below. Figure 2 As shown, the agent-based security verification device 30 includes: Processing module 301 is used to input the received user query into the security classification agent, and to perform security detection on the user query through the cascading detection mechanism of the security classification agent to obtain a security judgment result; Based on the security determination result, if the user query is an insecure query, a preset interception response is executed and a security event log is recorded; if the user query is a secure query, the user query is forwarded to the business processing intelligent agent. The business processing intelligent agent executes the financial case verification logic to obtain a business response; Output module 302 is used to input the business response to the output review agent, and the output review agent performs sensitivity checks and compliance checks on the business response to obtain the output review result; Based on the output review result, if the output review result is passed, the business response is output; if the output review result is not passed, the business response is replaced with a preset security response and then output, and all security events are recorded in the audit log module.

[0110] Optionally, the processing module 301 is further configured to: The user query is subjected to keyword matching detection by a rule filter. The rule filter contains a preset number of financial-specific malicious patterns. If the match is successful, the first security judgment result is output as unsafe and the first confidence level is output. If the first confidence score of the rule filter is lower than the preset first confidence score threshold, then the small model classifier is invoked to perform semantic analysis on the user query. The small model classifier is a Chinese pre-trained language model that has been supervised and fine-tuned by a preset amount of financial security labeled corpus, and outputs a second security judgment result and a second confidence score. If the second confidence score of the small model classifier is lower than the preset first confidence score threshold, the large language model classifier is invoked to perform deep semantic analysis on the user query. The large language model classifier identifies multiple threat types through preset prompt words containing security classification instructions and outputs a third security judgment result, threat type and third confidence score. Combining the first security assessment result, the second security assessment result, and the third security assessment result, if the security assessment result at any level is unsafe, the target security assessment result is unsafe, and the highest value among the first confidence level, the second confidence level, and the third confidence level is taken as the confidence level.

[0111] Optionally, the processing module 301 is further configured to: Generate a preset standard rejection message as the interception response, and return the standard rejection message to the user terminal; Obtain the threat type and confidence level output by the security classification agent, and record the security event log to the audit log module. The security event log includes the user identifier, the user query, the threat type, the confidence level, and the current timestamp. The subsequent process is terminated, the forwarding of the user query to the business processing agent is prohibited, and the execution of the financial case verification logic and the inspection of the output review agent are prohibited.

[0112] Optionally, the processing module 301 is further configured to: The intent of the user query is analyzed, and business parameters are extracted from the user query. The business parameters include case identifier and user information. Based on the business parameters, query the corresponding case information and user transaction records, match the case information with the preset risk control rules, and obtain the risk control rule matching result; Risk scores are calculated based on the risk control rule matching results and a preset risk scoring model, wherein the risk scoring model combines a rule engine and a machine learning model. A risk analysis report is generated based on the risk score and the case information, and the risk analysis report is output as the business response.

[0113] Optionally, the output module 302 is further configured to: The business response is subjected to a sensitivity check, which detects whether the business response contains sensitive information through rule matching and semantic analysis. The sensitive information includes system prompt words, source code, application programming interface keys, and risk control rule thresholds. The business response is subject to a compliance check, which examines whether the business response complies with financial regulatory requirements, whether it contains inappropriate advice or promises, and whether the data disclosure is compliant. If both the sensitivity check and the compliance check pass, the output review result is "passed". If either the sensitivity check or the compliance check fails, the output review result is "failed".

[0114] Optionally, the output module 302 is further configured to: Record the security event logs corresponding to all blocked user queries. The security event logs include user ID, query content, threat type, confidence level, and timestamp. Record the pass logs for all user queries that pass the security check. The pass logs include user identifier, query content, business response summary, and timestamp. Record output review event logs, which include output review results, sensitive information types, and replaced response content; The security event logs, the pass logs, and the output review event logs are stored in a traceable database, which supports querying by user ID, time range, and threat type.

[0115] Optionally, the output module 302 is further configured to: Record the security event logs corresponding to all blocked user queries. The security event logs include user ID, query content, threat type, confidence level, and timestamp. Record the pass logs for all user queries that pass the security check. The pass logs include user identifier, query content, business response summary, and timestamp. Record output review event logs, which include output review results, sensitive information types, and replaced response content; The security event logs, the pass logs, and the output review event logs are stored in a traceable database, which supports querying by user ID, time range, and threat type.

[0116] This application embodiment constructs a multi-agent collaborative architecture consisting of a security classification agent, a business processing agent, and an output review agent, and executes the entire process of cascading detection, security checks, business processing, output review, and audit log recording. From input to output, it provides comprehensive and multi-layered security protection for the financial case verification process, improving the attack interception rate, reducing the false alarm rate, and meeting the real-time, compliance, and traceability requirements of financial scenarios.

[0117] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided in this application can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0118] It should be noted that, in the embodiments of this application, if the above-described security verification method based on intelligent agents is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0119] This application provides a computer device including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method.

[0120] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described method. The computer-readable storage medium can be transient or non-transient.

[0121] This application provides a computer program including computer-readable code, wherein when the computer-readable code is executed in a computer device, a processor in the computer device performs some or all of the steps in the above-described method.

[0122] This application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described method. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium; in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc.

[0123] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referred to interchangeably. The descriptions of the above embodiments of the device, storage medium, computer program, and computer program product are similar to the descriptions of the above method embodiments and have similar beneficial effects. For technical details not disclosed in the embodiments of the device, storage medium, computer program, and computer program product of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0124] It should be noted that, Figure 3 This is a schematic diagram of a hardware entity of a computer device in an embodiment of this application, such as... Figure 3 As shown, the hardware entity of the computer device 700 includes: one or more processors 701, a communication interface 702, and a memory 703, wherein: Processor 701 typically controls the overall operation of computer device 700.

[0125] Communication interface 702 enables computer devices to communicate with other terminals or servers over a network.

[0126] The memory 703 is configured to store instructions and applications executable by the processor 701, and can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) in the processor 701 and various modules in the computer device 700. It can be implemented using flash memory or random access memory (RAM). Data transfer between the processor 701, the communication interface 702, and the memory 703 can be performed via bus 704. Only one processor is shown in the figure; each processor 701 includes one or more cores.

[0127] It should be noted that the computer device may include multiple processors 701, and each processor 701 can interact with each other through aggregated communication methods such as all-to-all, all-gather, or all-reduce. The processors 701 may be central processing units (CPUs), graphics processing units (GPUs), embedded neural network processing units (NPUs), tensor processing units (TPUs), data processing units (DPUs), accelerated processing units (APUs), floating-point processing units (FPUs), or application-specific integrated circuits (ASICs). The processors may also be single-core or multi-core processors. The processor may consist of a CPU and hardware chips. The hardware chips may be ASICs, PLDs, or combinations thereof. The PLDs may be complex programmable logic devices (CPLDs), FPGAs, generic array logic (GALs), or any combination thereof. The processor can also be implemented using logic devices with built-in processing logic, such as FPGAs or digital signal processors (DSPs).

[0128] The communication interface 702 can be a wired interface or a wireless interface, used to communicate with other modules or devices. The wired interface can be an Ethernet interface, a local interconnect network (LIN), etc., and the wireless interface can be a cellular network interface or a wireless LAN interface, etc.

[0129] Memory 703 can be non-volatile memory, such as read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Memory 703 can also be volatile memory, which can be random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct rambus RAM (DRRAM), direct rambus DRAM (DRDRAM), and rambus DRAM.

[0130] The 704 bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0131] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above steps / processes do not imply a sequential order of execution; the execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above embodiments of this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0132] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0133] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0134] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0135] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0136] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0137] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, magnetic disks, or optical disks.

[0138] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A security verification method based on intelligent agents, characterized in that, The method includes: The received user query is input into the security classification agent, and the security of the user query is detected by the cascaded detection mechanism of the security classification agent to obtain a security judgment result. Based on the security determination result, if the user query is an insecure query, a preset interception response is executed and a security event log is recorded; if the user query is a secure query, the user query is forwarded to the business processing intelligent agent. The business processing intelligent agent executes the financial case verification logic to obtain a business response; The business response is input to the output review agent, which performs sensitivity and compliance checks on the business response to obtain the output review result. Based on the output review result, if the output review result is passed, the business response is output; if the output review result is not passed, the business response is replaced with a preset security response and then output, and all security events are recorded in the audit log module.

2. The method according to claim 1, characterized in that, The security detection of the user query through the cascading detection mechanism of the security classification agent includes: The user query is subjected to keyword matching detection by a rule filter. The rule filter contains a preset number of financial-specific malicious patterns. If the match is successful, the first security judgment result is output as unsafe and the first confidence level is output. If the first confidence score of the rule filter is lower than the preset first confidence score threshold, then the small model classifier is invoked to perform semantic analysis on the user query. The small model classifier is a Chinese pre-trained language model that has been supervised and fine-tuned by a preset amount of financial security labeled corpus, and outputs a second security judgment result and a second confidence score. If the second confidence score of the small model classifier is lower than the preset first confidence score threshold, the large language model classifier is invoked to perform deep semantic analysis on the user query. The large language model classifier identifies multiple threat types through preset prompt words containing security classification instructions and outputs a third security judgment result, threat type and third confidence score. Combining the first security assessment result, the second security assessment result, and the third security assessment result, if the security assessment result at any level is unsafe, the target security assessment result is unsafe, and the highest value among the first confidence level, the second confidence level, and the third confidence level is taken as the confidence level.

3. The method according to claim 1, characterized in that, The execution of the preset interception response and recording of security event logs includes: Generate a preset standard rejection message as the interception response, and return the standard rejection message to the user terminal; Obtain the threat type and confidence level output by the security classification agent, and record the security event log to the audit log module. The security event log includes the user identifier, the user query, the threat type, the confidence level, and the current timestamp. The subsequent process is terminated, the forwarding of the user query to the business processing agent is prohibited, and the execution of the financial case verification logic and the inspection of the output review agent are prohibited.

4. The method according to claim 1, characterized in that, The process of obtaining a business response by executing financial case verification logic through the business processing intelligent agent includes: The intent of the user query is analyzed, and business parameters are extracted from the user query. The business parameters include case identifier and user information. Based on the business parameters, query the corresponding case information and user transaction records, match the case information with the preset risk control rules, and obtain the risk control rule matching result; Risk scores are calculated based on the risk control rule matching results and a preset risk scoring model, wherein the risk scoring model combines a rule engine and a machine learning model. A risk analysis report is generated based on the risk score and the case information, and the risk analysis report is output as the business response.

5. The method according to claim 1, characterized in that, The process of performing sensitivity and compliance checks on the business response through the output review agent to obtain the output review results includes: The business response is subjected to a sensitivity check, which detects whether the business response contains sensitive information through rule matching and semantic analysis. The sensitive information includes system prompt words, source code, application programming interface keys, and risk control rule thresholds. The business response is subject to a compliance check, which examines whether the business response complies with financial regulatory requirements, whether it contains inappropriate advice or promises, and whether the data disclosure is compliant. If both the sensitivity check and the compliance check pass, the output review result is "passed". If either the sensitivity check or the compliance check fails, the output review result is "failed".

6. The method according to claim 1, characterized in that, The method of recording all security events to the audit log module includes: Record the security event logs corresponding to all blocked user queries. The security event logs include user ID, query content, threat type, confidence level, and timestamp. Record the pass logs for all user queries that pass the security check. The pass logs include user identifier, query content, business response summary, and timestamp. Record output review event logs, which include output review results, sensitive information types, and replaced response content; The security event logs, the pass logs, and the output review event logs are stored in a traceable database, which supports querying by user ID, time range, and threat type.

7. The method according to claim 4, characterized in that, The method further includes: Record the security event logs corresponding to all blocked user queries. The security event logs include user ID, query content, threat type, confidence level, and timestamp. Record the pass logs for all user queries that pass the security check. The pass logs include user identifier, query content, business response summary, and timestamp. Record output review event logs, which include output review results, sensitive information types, and replaced response content; The security event logs, the pass logs, and the output review event logs are stored in a traceable database, which supports querying by user ID, time range, and threat type.

8. A security verification device based on an intelligent agent, characterized in that, The device includes: The processing module is used to input the received user query into the security classification agent, and to perform security detection on the user query through the cascading detection mechanism of the security classification agent to obtain a security judgment result; Based on the security determination result, if the user query is an insecure query, a preset interception response is executed and a security event log is recorded; if the user query is a secure query, the user query is forwarded to the business processing intelligent agent. The business processing intelligent agent executes the financial case verification logic to obtain a business response; The output module is used to input the business response to the output review agent, and the output review agent performs sensitivity checks and compliance checks on the business response to obtain the output review result. Based on the output review result, if the output review result is passed, the business response is output; if the output review result is not passed, the business response is replaced with a preset security response and then output, and all security events are recorded in the audit log module.

9. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the agent-based security verification method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the agent-based security verification method according to any one of claims 1 to 7.

Images