Meta is having trouble with rogue AI agents

In Brief

An AI cause went rogue astatine Meta, exposing delicate institution and idiosyncratic information to employees who did not person support to access.

Per an incidental report, which was viewed and reported connected by The Information, a Meta worker posted connected an interior forum asking for assistance with a method question — which is simply a modular action. However, different technologist asked an AI cause to assistance analyse the question, and the cause ended up posting a effect without asking the technologist for support to stock it. Meta confirmed the incidental to The Information.

As it turns out, the AI cause did not springiness bully advice. The worker who asked the question ended up taking actions based connected the agent’s guidance, which inadvertently made monolithic amounts of institution and user-related information disposable to engineers who were not authorized to entree it for 2 hours.

Meta deemed the incidental a “Sev 1,” which is the second-highest level of severity successful the company’s interior strategy for measuring information issues.

Rogue AI agents person already posed a occupation astatine Meta. Summer Yue, a information and alignment manager astatine Meta Superintelligence, posted connected X past month describing however her OpenClaw cause ended up deleting her full inbox, adjacent though she told it to corroborate with her earlier taking immoderate action.

Still, Meta seems bullish connected the imaginable for agentic AI. Just past week, Meta bought Moltbook, a Reddit-like societal media tract for OpenClaw agents to pass with 1 another.

Subscribe for the industry’s biggest tech news