Nobody measures how long an agent keeps working after you revoke its access
anp2_protocol · reddit · 2026-08-21
A deep Reddit thread on a blind spot in agent security: tstop — the elapsed time from revoking credentials to the last successful privileged call.
- Revocation is a write on the issuing side, but cached tokens with remaining TTL, open sessions, queued jobs carrying credentials, sub-agents with copied credentials, and in-flight tool calls keep working afterward.
- Most teams only verify the revoke API returns 200 and have never measured tstop — which can be measured in an afternoon: run a loop, revoke mid-run, find the timestamp of the last successful call.
- Shorter TTLs make the refresh path the real authority, and rules like "renew while the job is healthy" can re-grant credentials during exactly the incident you're trying to stop.
- Shrinking tstop means per-call authorization checks: fail open and you revoked nothing; fail closed and your authorizer becomes an outage amplifier.
- Least measured: every side effect has its own tstop — queued messages and payment records with idempotency keys still fire after access is gone, and "revoke to last externally visible effect" spans two systems nobody owns.
More from coding & agent
- MiniMax Code CLI Integrates into CI to Auto-Fix Build Failures — HeyAmit_ · 2026-08-21
- NIO Adopts Volcano Engine TRAE, Achieving 90% AI Code Adoption Rate — 火山引擎 · 2026-08-21
- Grok Bot hailed as the best AI agent, video showcases 8 use cases — PeterDiamandis · 2026-08-21
- Storing Agent Outputs with Permissions and Audit Trails — RocketSeven · 2026-08-21
- Prompt vs context engineering misses 3 layers: the 5-layer stack for shipping AI systems — blaizedsouza · 2026-08-21
- How to give every coding agent the same second brain — blaizedsouza · 2026-08-21