The audit log says my name: what an agent inherits when you hand it your credentials
I measured what an AI agent actually holds when it authenticates as me. Authentication turned out to be a token The post The audit log says my name: what an agent inherits when you hand it your credentials appeared first on The New Stack .
The audit log reveals the extent of what an AI agent inherits when you provide it with an individual's credentials. When the agent authenticates as the user, it utilizes a token with a scope string explicitly stating "user_impersonation". Authorization primarily stems from group membership, which standard permission checks cannot detect.
Permissions differ significantly between two production stores, with one allowing 107 permissions, while the other has only two. The environment where the agent could delete from a secure database is the sole one lacking an audit trail. When using borrowed credentials, three common outcomes arise: attribution becomes unclear, the agent inherits the human's accumulated permissions, and revoking the agent also revokes the human's access.
The agent's actions are virtually indistinguishable from the human's, as the database records do not differentiate based on whether the query was executed by the agent or the user. Authentication is straightforward, with no password storage or separate identity provisioning required. The agent piggybacks on the cached token from the human's earlier authentication session.
The access token reveals that the agent operates under the "user_impersonation" scope and includes an "mfa" authentication method, despite the 84-minute lifetime for the token being insufficient for measuring intent. Group membership plays a crucial role, with 37 entries in the token's claim versus 34 visible subscriptions in the directory.
Some memberships are managed through just-in-time (JIT) processes, and the database uses the higher count for authority checks.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
