AI agent security

Review what an agent can access, what its tools can do, and which information can appear in its outputs.

Questions for a security review

Does the model need the sensitive value itself, or can an authorized tool use it without returning it?
Which permissions limit the tools available to the agent?
Can untrusted input lead to an unauthorized action?
What checks apply before information leaves the system?
What records are available to investigate a failed control?

Controls to examine

Our work examines protected values accessed through approved tools, signed policy records, and checks on outgoing data. We assess each control against the application, its tool behavior, and the ways data can leave it.

TN-Proto is our research and development work on permissioned event records. Its documentation describes the protocol and implementation.

Related guidance

See data governance for access and record-handling questions, and audit trails for the evidence needed to review actions.