← sources

How are you handling revocation for AI agents that sit idle for months?

source: reddit (reddit:sysadmin)published 2026-10-31 17:10ingested 2026-11-01 02:39

retained text

This is the dormant account problem, just with agents now. The Arrakis report put it at 76% of active agents sitting idle 60+ days with their credentials still working, and 24% startable by an unsigned webhook.

With a human account, you disable it and move on. With an agent, killing the identity doesn't always kill the access — there's usually a cached token, a shared service key, or a webhook that can start it back up somewhere else.

The way I've been approaching it is boring on purpose: approvals expire on their own, nothing lives forever, and every decision lands in a proof log so I can show no auth events after the retire date, not just say it was revoked.

How do you handle this in your shops? Do you treat agent credentials like service accounts, give them short TTLs, or is there something else that actually works when an agent has been idle for months?

document fetched 2026-11-01 00:33 · canonical https://reddit.com/r/sysadmin/comments/1xkfnf9/how_are_you_handling_revocation_for_ai_agents