Researchers have proposed a four-layer security architecture designed to give artificial intelligence agents persistent identities, verifiable capability records, bounded authorization and cryptographic accountability when they interact across organizations.
The proposed protocol suite, called InterSAGE, addresses a gap the authors see in emerging agent communication standards. Protocols can define how agents discover peers, exchange messages and invoke tools, but they do not necessarily establish who controls an agent, what software it is running or whether its claimed permissions can be trusted.
InterSAGE begins with an Agent Identity Card that binds four elements into one verifiable credential: the developer, code package, operator and deployment context. A Global Agent Registry would sign the credential and support key provisioning, rotation and revocation.
The discovery layer would represent an agent’s skills and tools as Verifiable Credentials linked to its decentralized identifier. Before an interaction begins, another agent could check the issuer, confirm that the credential belongs to the presenting agent and determine whether the advertised capability fits its permissions.
A trust-negotiation layer would use mutual challenge-response checks and calculate a session’s permissions by intersecting the authority granted at different stages. The design is intended to ensure that delegated capabilities can only be narrowed as they move from a developer to an operator and then to a running agent.
The final layer would create kernel-mediated cryptographic audit trails that connect tool use, delegation and execution records to an agent’s identity. The architecture does not require a shared blockchain, though deployments could add hardware security modules, trusted execution environments or other trust anchors.
The authors compare InterSAGE with more than 50 projects spanning agent protocols, decentralized identity, OAuth extensions, zero-trust systems and audit frameworks. They characterize the work as a positioning paper rather than a production-ready specification; detailed proofs, performance benchmarks and implementation results are reserved for later work.
The proposal arrives as public bodies begin defining identity rules for autonomous software. NIST has opened work on AI agent identity and authorization, while the ITU is examining identity and trust for humans and AI agents.
InterSAGE is designed to complement rather than replace communication protocols such as MCP, A2A, ANP and AG-UI. Its contribution is a common trust layer intended to carry identity, capability and accountability information as agents move between different technical environments.
Sources: arXiv
–
By the ID Tech Editorial Team
