AgentTrust 1.0 puts a human approval gate in front of agent tools
The October 25 release combines capability policies, approval receipts and audit records, with security assurances still requiring evaluation.
AgentTrust 1.0.0 proposes a control point between an AI agent and the tools it wants to execute. Published October 25, its release notes describe deny-by-default capability policies, narrower permissions for delegated agents, human approval for sensitive actions and an audit ledger.
The useful workflow is concrete: an agent proposes an action, policy checks decide whether it is allowed, and higher-impact work can require a person to approve it. The authors describe single-use approval nonces intended to prevent reuse of an authorization. They also advertise outbound-request filtering and spending boundaries.
Those features should not be confused with a demonstrated guarantee against prompt injection or data theft. The release uses strong security and post-quantum language, but naming a cryptographic method does not establish that an entire deployment is secure or compliant. AITrending has not audited the implementation or reproduced the protection claims.
The release lists a versioned npm package and a packed distribution asset. The project also documents running its server and dashboard locally after installing dependencies and building the client. It is a gateway around agent execution, not a model provider or an included inference plan. Evaluate the October 25 release as a dated version; the repository’s current README can describe later work.