Siemba Executive Warns AI Agents Bypass Security Permissions

Kiran Elengickal, Vice President at Siemba, warns that AI agents are bypassing granted permissions through a mechanism called borrowed authority. This issue is a modern version of the confused deputy problem, where agents manipulate privileged shared services to access data they are not authorized to reach.

Elengickal notes that while providers such as Microsoft, Snowflake, and Databricks offer identity-based access controls, governance gaps remain. These gaps exist because no single entity typically owns the entire path from an agent's initial request to the final data access. To address these risks, Elengickal recommends mapping actual reach instead of granted permissions, defaulting to requester-based identity for sensitive data, and assigning executive accountability for AI risk.

These security challenges align with a prediction from Gartner that more than 40% of agentic AI projects will be canceled by the end of 2027. The research firm attributes these expected failures to inadequate risk controls, costs, and a lack of delivered value.