Cyber Risks · The Hacker News

AI Agents Are Becoming Authorization Bypass Paths

Summary: Organizational AI agents authenticate through shared service accounts, API keys, or OAuth grants that are often long-lived and granted broad permissions so a single agent can serve many users and workflows. Because actions execute under the agent's identity rather than the requester's, a user with limited access can retrieve data they could not access directly, and logs attribute the activity to the agent, not the person who asked. The article illustrates this with a roughly 1,000-employee technology and marketing solutions company where a new hire with intentionally limited permissions asked a Databricks-connected marketing agent to analyze churn and received detailed sensitive customer data, with nothing misconfigured and no policy violated.

Why it matters: IAM enforces permissions against the agent's identity, not the requester's, so least privilege no longer holds and the audit trail names the agent instead of the person who initiated the action. In the article's example, a new hire with intentionally limited permissions received detailed sensitive customer data through an agent that was working exactly as designed.

Read the original at The Hacker News ↗