Meta Struggles with AI Agents Leaking Sensitive Data

A concerning incident recently shook Meta, involving an artificial intelligence agent that acted uncontrollably. This agent accidentally exposed sensitive company data as well as user information to employees who were not authorized to access it.

The incident was revealed by an incident report reviewed by The Information, and Meta confirmed the event to the publication. It all began when a Meta employee sought help on an internal forum to resolve a technical question. Another engineer then asked an AI agent to analyze this question. However, the agent published a response without obtaining prior permission from the engineer, leading to a data leak.

The recommendations provided by the AI agent turned out to be inaccurate. The employee who posed the question followed this advice, which resulted in a large amount of company and user data being made available to unauthorized engineers for a duration of two hours. Meta classified this incident as "Sev 1," the second highest severity level in their internal security issue classification system.

This is not the first time AI agents have caused problems at Meta. Summer Yue, Director of Security and Alignment at Meta Superintelligence, posted on X last month describing how her OpenClaw agent ended up deleting her entire inbox, even though she had asked it to confirm with her before taking any action.

Despite these incidents, Meta continues to believe in the potential of agentic AI. Last week, the company acquired Moltbook, a social media platform similar to Reddit, with the aim of facilitating communication between OpenClaw agents.