A cross-platform AI agent permission negotiation method

Abstract

The application discloses a cross-platform AI agent permission negotiation method and relates to the technical field of artificial intelligence agent security. The method realizes two-way identity verification through a standardized handshake message, aligns heterogeneous permission models by adopting a unified permission dimension field, generates mutual recognition authorization snapshots through a two-way negotiation process of declaration by an initiating end and clipping by a receiving end, and maintains consistency of authorization states of two ends through an independent state synchronization channel. The application solves the technical problems that permissions between heterogeneous AI agents of different manufacturers and different systems cannot be mutually recognized and authorization states are inconsistent, and can be applied to a multi-agent cross-platform interconnection scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence agent security technology, and in particular to a method for permission negotiation and state synchronization between AI agents deployed in different computing systems. Background Technology

[0002] With the development of AI agent technology, AI agents deployed within a single system can no longer meet complex business needs. AI agents from different vendors and systems need to collaborate across platforms. Currently, AI agent access control solutions are all deployed within a single computing system, operating based on a unified account system and a unified access control model.

[0003] When AI agents belonging to different systems interact across platforms, the following technical shortcomings exist: First, the dimensional definitions of the permission models on both ends are inconsistent, making it impossible to directly identify the other party's permission statements; second, there is a lack of standardized two-way identity verification processes during cross-domain interactions; third, after authorization, the permission states on both ends are maintained independently, which can easily lead to inconsistencies. Therefore, a technical solution is needed that can achieve standardized permission negotiation between heterogeneous AI agents. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a cross-platform AI agent permission negotiation method to address the technical issues of heterogeneous AI agent permission models, lack of cross-domain identity verification, and inconsistent authorization status in different computing systems. This method achieves trusted permission mutual recognition between heterogeneous AI agents through standardized message structure, unified permission dimensions, bidirectional negotiation process, and status synchronization mechanism.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The first AI agent and the second AI agent are deployed in independent computing systems, each maintaining its local permission model. The first AI agent generates and sends a handshake request message containing a domain identifier, identity credentials, and permission metadata template. The second AI agent receives the request, performs identity verification, and returns a handshake response message containing the corresponding fields. After completing bidirectional verification, the first AI agent generates and sends an authorization permission set containing three dimensions: operation type, resource scope, and timeliness. The second AI agent performs boundary trimming on the three dimensions according to its local security policy, generates a trimmed permission set, and returns it. Both parties store the trimmed permission sets as mutual authorization snapshots. Both parties establish a state synchronization channel; when the authorization state changes at either end, it notifies the other end via a state change message and updates synchronously.

[0006] Furthermore, the permission metadata template includes operation type enumeration values, resource type enumeration values, and time-limited unit enumeration values, which are used for mapping and alignment of permission dimensions at both ends.

[0007] Furthermore, boundary pruning includes pruning processing in three dimensions: comparison of resource scope with local resource whitelist, comparison of operation type with local operation permission set, and comparison of timeliness with local maximum authorization duration.

[0008] Furthermore, when a mutually recognized authorized snapshot is generated, both parties perform hash calculations on the snapshot content and exchange hash values. When the hash values ​​match, the negotiation is confirmed to be complete.

[0009] Furthermore, the authorization status change includes four statuses: authorization effective, authorization frozen, permission downgraded, and authorization revoked. The status change message includes the identifier of the entity performing the change and the change timestamp.

[0010] Furthermore, the various messages during the interaction process are stored in a structured manner to generate an authorized link log containing message hashes, timestamps, and domain identifiers of both parties.

[0011] The beneficial effects of this invention are as follows: it achieves dimensional alignment between different permission models through a unified permission metadata template; it achieves cross-domain identity verification through bidirectional handshake messages; it achieves mutually agreed permission boundaries through a send-truncate-confirm negotiation process; and it ensures consistency of authorization states at both ends through an independent state synchronization channel. These technical features together constitute a complete cross-domain permission negotiation mechanism, capable of supporting the secure operation of multi-agent cross-platform interconnection. Detailed Implementation

[0012] This embodiment provides a cross-platform AI agent permission negotiation method. A first AI agent is deployed on a first computing system, and a second AI agent is deployed on a second computing system. The two computing systems are independent of each other and each maintains an independent account system and permission model. The first and second computing systems interact with each other via a network connection.

[0013] This method includes the following steps: Step 1, Trusted Handshake Phase. The first AI agent generates a handshake request message. The message structure includes a header, a first domain identifier field, a first identity credential field, a first permission metadata template field, and a trailer. The first domain identifier field identifies the domain to which the first computing system belongs; the first identity credential field verifies the legitimacy of the first AI agent's identity; and the first permission metadata template field contains a set of enumerated values ​​for operation types, resource types, and time units. The first AI agent sends the handshake request message to the second AI agent over the network.

[0014] After receiving the handshake request message, the second AI agent extracts the first domain identifier field and the first identity credential field, and compares them with the locally stored list of valid domains and the identity credential database. If the verification fails, the interaction terminates; if the verification succeeds, a handshake response message is generated. The message structure includes a message header, a second domain identifier field, a second identity credential field, a second permission metadata template field, and a message trailer. The second AI agent then returns the handshake response message to the first AI agent.

[0015] After receiving the handshake response message, the first AI agent extracts the second domain identifier field and the second identity credential field, and compares them with the locally stored list of legitimate domains and the identity credential database. If the verification fails, the interaction terminates; if the verification succeeds, a two-way trusted identity handshake is completed, and the authorization negotiation phase begins.

[0016] Step Two: Permission Negotiation Phase. The first AI agent generates a set of authorized permissions based on its local permission policy. This set is represented using structured fields, including an operation type dimension field, a resource scope dimension field, and a timeliness dimension field. Specifically, the operation type dimension field takes a subset of the operation type enumeration values ​​in the first permission metadata template, the resource scope dimension field takes a subset of the resource type enumeration values ​​in the first permission metadata template, and the timeliness dimension field takes the start and end timestamps. The first AI agent then sends the authorized permission set to the second AI agent.

[0017] After receiving the authorized permission set, the second AI agent performs boundary pruning based on the local security policy. The pruning process consists of three dimensions: First, resource scope pruning, comparing the resource scope field in the authorized permission set with the local resource whitelist and removing resources not in the whitelist; second, operation type pruning, comparing the operation type field in the authorized permission set with the local operation permission set and removing operations not in the permission set; third, timeliness pruning, comparing the timeliness field in the authorized permission set with the local maximum authorized duration, and if the termination timestamp exceeds the local maximum authorized duration, pruning the termination timestamp to the time point corresponding to the maximum authorized duration.

[0018] After the second AI agent completes the three-dimensional cropping, it generates a cropped permission set and returns the cropped permission set to the first AI agent.

[0019] Step 3, Mutual Recognition Snapshot Generation Phase. The first AI agent receives the trimmed permission set, and the second AI agent locally retains the trimmed permission set. Both parties store the trimmed permission set as a mutual recognition authorization snapshot. The data structure of the mutual recognition authorization snapshot includes: first domain identifier, second domain identifier, negotiation timestamp, operation type dimension field, resource scope dimension field, timeliness dimension field, and unique authorization identifier.

[0020] When a snapshot is generated, the first AI agent performs a hash calculation on the locally recognized authorized snapshot content, generates a first hash value, and sends it to the second AI agent; the second AI agent performs a hash calculation on the locally recognized authorized snapshot content, generates a second hash value, and sends it to the first AI agent. Both parties compare the received hash value with the locally calculated hash value. If they match, the negotiation is confirmed to be complete, and the mutually recognized authorized snapshot officially takes effect.

[0021] Step four, State Synchronization Phase. A long-lived state synchronization channel is established between the first AI agent and the second AI agent. When an authorization state change occurs at either end, a state change message is generated. The message structure includes a unique authorization identifier, a status code, a change operation subject identifier, and a change timestamp. The status code can have four values: authorization effective, authorization frozen, permission downgraded, and authorization revoked. The state change message is sent to the other end through the state synchronization channel. Upon receiving the message, the other end locates the corresponding mutually recognized authorization snapshot based on the unique authorization identifier and updates the status field in the snapshot.

[0022] Step 5, Link Evidence Storage Stage. The handshake request messages, handshake response messages, authorization permission sets, trimmed permission sets, and state change messages generated during the interaction process are stored in a structured manner. Each stored record includes a message hash value, a message reception timestamp, a sender domain identifier, and a receiver domain identifier, forming a complete authorization link log for subsequent auditing and tracing.

Claims

1. A cross-platform AI agent permission negotiation method, characterized in that, Includes the following steps: The first AI agent and the second AI agent are deployed in independent computing systems, and each of the first AI agent and the second AI agent maintains its own local permission model. S1. The first AI agent generates a handshake request message, which includes a first domain identifier field, a first identity credential field, and a first permission metadata template field, and sends the handshake request message to the second AI agent. S2. The second AI agent receives the handshake request message, extracts the first domain identifier field and the first identity credential field for legality verification, and generates a handshake response message after the verification is passed. The handshake response message includes the second domain identifier field, the second identity credential field, and the second permission metadata template field. The handshake response message is then returned to the first AI agent. S3. The first AI agent receives the handshake response message, extracts the second domain identifier field and the second identity credential field for legality verification, and generates an authorization permission set after the verification is passed. The authorization permission set includes an operation type dimension field, a resource scope dimension field, and a timeliness dimension field. The authorization permission set is then sent to the second AI agent. S4. The second AI agent receives the authorized permission set, performs boundary trimming on the operation type dimension field, resource scope dimension field, and timeliness dimension field according to the local security policy, generates a trimmed permission set, and returns the trimmed permission set to the first AI agent. S5. The first AI agent and the second AI agent respectively store the trimmed permission set as a mutual authorization snapshot. The mutual authorization snapshot includes the domain identifiers of both parties, the negotiation timestamp, the permission dimension field, and the unique authorization identifier. S6. A state synchronization channel is established between the first AI agent and the second AI agent. When an authorization state change occurs at either end, a state change message containing the unique authorization identifier and status code is generated and sent to the other end through the state synchronization channel. After receiving the message, the other end updates the corresponding state of its local mutual authorization snapshot.

2. The method according to claim 1, characterized in that, Both the first permission metadata template field and the second permission metadata template field contain operation type enumeration values, resource type enumeration values, and time-effect unit enumeration values, which are used for mapping and aligning the permission dimensions at both ends.

3. The method according to claim 1, characterized in that, The boundary trimming in S4 specifically includes: comparing the resource range dimension field with the local resource whitelist and removing resource items that are not in the whitelist; comparing the operation type dimension field with the local operation permission set and removing operation items that are not in the permission set; and comparing the timeliness dimension field with the local maximum authorization duration and trimming any excess to the maximum authorization duration.

4. The method according to claim 1, characterized in that, In step S5, when the mutual recognition and authorization snapshot is generated, the first AI agent and the second AI agent respectively perform hash calculation on the snapshot content and exchange hash values. When the hash values ​​are consistent, the negotiation is confirmed to be completed.

5. The method according to claim 1, characterized in that, The status codes in S6 include four types: authorization effective, authorization frozen, permission downgraded, and authorization revoked. The status change message includes the change operation subject identifier and the change timestamp.

6. The method according to claim 1, characterized in that, It also includes S7: structured storage of handshake request messages, authorization permission sets, trimmed permission sets, and state change messages, generating authorization link logs containing message hashes, timestamps, and domain identifiers of both parties.