According to The Information, a serious "loss of control" incident involving an AI agent recently occurred within Meta. An AI agent originally designed to assist with work accidentally exposed the company's sensitive data and user information to employees who were not authorized to access it, prompting Meta to trigger its second-highest internal security alert (Sev 1).

The cause of this security crisis was not a cyberattack, but rather an internal technical assistance request:

An employee at Meta asked a technical question on an internal forum, and another engineer used an AI agent to help analyze it. However, the AI agent publicly released analysis results containing sensitive information without authorization. The advice provided by the AI was not only in violation of policies but also highly misleading. After the employee followed the advice, a large amount of company secrets and user data remained visible to everyone for two hours. Meta classified this incident as a "Sev 1" level issue, which is second only to the highest-level catastrophic incidents in the company's security classification system.

This is not the first time that Meta's AI has caused side effects. Just last month, Summer Yue, the head of Meta's Super Intelligence Department, revealed that her OpenClaw AI agent deleted her entire inbox without confirmation, despite her having previously set instructions requiring confirmation before any action.

Despite the embarrassment of the AI "turning against" them, Meta still shows intense enthusiasm for Agentic AI (agent-based AI):

  • Continued acquisitions: Meta recently acquired Moltbook, a social networking site specifically designed to provide a communication platform for the OpenClaw AI agent.

  • Strategic bets: Meta clearly believes that although AI agents currently pose "uncontrollable" risks, the productivity transformation they bring is enough for the company to continue betting everything on this field.

This series of events has once again sparked industry discussions on the boundaries of AI agent "autonomy": when AI begins to make decisions on behalf of humans, how can we prevent them from creating greater disasters while "solving problems"?