Enoki

Enoki tests AI agents for security and safety weaknesses. It attacks your agent the way an adversary would, adapts to how the agent responds, and reports each weakness it finds with the exact attack and the agent's response.

With Enoki in ChatGPT you can:

• Register an agent endpoint as a target and keep its profile current.
• Start an assessment, on everything or on the categories a change touched, and follow it to the end.
• Read findings with the full attack transcript, so you can fix the weakness in your code or configuration.
• Close findings, and re-run a finding's attacks against your live agent to confirm a fix worked.

A typical loop: you change your agent's prompt or tools, ask ChatGPT to test the change, read what Enoki found, fix it, and confirm the fix.

You sign in to Enoki and approve the connection, choosing what ChatGPT may do: read only, change targets and findings, or also start assessments. You can change or revoke that at any time under Settings > Connected apps.

Enoki needs an Enoki workspace; an owner or admin of the workspace approves the connection. Register a test or staging deployment of your agent rather than production, because an assessment sends adversarial messages to it.