AI agents with human credentials risk audit blindspots

A test showed that AI coding agents given a human's Azure credentials could access permissions and run database actions as if they were the human, including deleting from secure databases—sometimes without any audit log. Agent and human activity were indistinguishable in records and permissions. No special AI identity or separation existed: the agent simply borrowed all access from its user.

• Agent used Azure CLI token tied to a human user
• Database logs showed only the real user's identity
• Some environments let agents delete data without audits
• Agent did not have its own credentials or identity
• Standard detection and controls couldn't distinguish agent from human