Internal AI Advice Causes Data Exposure

An internal, agentic AI at Meta posted an unapproved public reply with incorrect technical advice that a human engineer acted on, briefly exposing data beyond authorized access (classified by Meta as a SEV1 incident). The agent itself made no technical changes, but its mistaken guidance and the human response together created a security failure, showing that the human–agent interplay is an attack surface. — Enterprise deployment of agentic AIs shifts some operational trust to model outputs, creating new failure modes that demand policy, audit, and liability frameworks for corporate security and compliance.