How the attack works
The attacker sends the agent a request that looks like an ordinary query rather than an obvious cross-user lookup. The phrasing avoids naming another user explicitly but instead references an object indirectly, such as by balance size, transaction count, or a specific account number. If the agent's backend does not verify that the requester actually owns the record being fetched, it returns another user's data. This differs from role-based bypass attempts because it targets a specific object reference rather than a permission level.
Why it matters
An organization can leak individual users' financial or personal records to unauthorized requesters, one account at a time, without any obvious authorization error being triggered.
What you can do
- Enforce object-level ownership checks on every data fetch, not just role checks.
- Reject or flag agent queries that reference records by attributes (balance, count, ID) instead of the requester's own identity.
- Log and review cases where an agent fetches records not tied to the authenticated user's ID.
- Exclude known legitimate cases, such as analytics agents or admins with explicit cross-account read rights, from alerting to reduce noise.
Known benign look-alikes
- Authorized analytics agents querying aggregate data across all users
- Admin users with explicit cross-account read permissions