A permission review cycle for AI agents that teams will actually run
BlitzBallBruiser · reddit · 2026-11-17
A practical guide arguing that "review everything" permission audits stall, and proposing a cycle sized at 1 hour/month:
- Generate inventory: one row per grant — agent, reach (API scope/MCP server/tool), credential, date added — pulled from live grants in your IdP, secrets manager, and gateway config
- One human owner per row: team aliases can't answer "what is this for?"; ownerless rows get reviewed first
- Triage before review: sort by risk — write/delete/send/spend over read-only, production over staging; MCP hints like destructiveHint are untrusted per spec, so verify actual behavior
- Review only the riskiest half each cycle
- Compare granted vs needed: unused ≠ unneeded for quarterly/incident-only jobs; the owner decides
- Expire by default: owners pick expiry at grant time, shorter for riskier access; renewal shrinks the next cycle
- Keep a decision log: keep/narrow/remove, who, date, one-line reason, next review — what auditors ask for; unchanged recent "keep" rows can be skipped
Cites OWASP: excessive permissions are a root cause of Excessive Agency in LLM apps.
More from coding & agent
- Dev calls for an 'omni-feed' product: one filterable pane for all agent actions — DavidWells · 2026-11-17
- macOS 27.3 beta adds granular Full Disk Access controls, making Macs more agent-friendly — film_girl · 2026-11-17
- LangChain adds host allowlists to Managed Deep Agents, blocking injected exfil via proxy — hwchase17 · 2026-11-17
- Goodhart's Law meets AI agents: stop counting PRs, measure the delivery system — rseroter · 2026-11-17
- Running 10 AI agents in parallel: "I'm the bottleneck of my own company" — stewones · 2026-11-17
- Bitter lesson comes for multi-agent systems: RL-trained swarms beat hand-designed orchestration — emax · 2026-11-17