An AI agent went rogue at Meta, exposing delicate firm and consumer knowledge to workers who didn’t have permission to entry it.
Per an incident report, which was considered and reported on by The Information, a Meta worker posted on an inner discussion board asking for assist with a technical query — which is a typical motion. Nevertheless, one other engineer requested an AI agent to assist analyze the query, and the agent ended up posting a response with out asking the engineer for permission to share it. Meta confirmed the incident to The Data.
Because it seems, the AI agent didn’t give good recommendation. The worker who requested the query ended up taking actions based mostly on the agent’s steering, which inadvertently made large quantities of firm and user-related knowledge out there to engineers, who weren’t approved to entry it, for 2 hours.
Meta deemed the incident a “Sev 1,” which is the second-highest degree of severity within the firm’s inner system for measuring safety points.
Rogue AI brokers have already posed an issue at Meta. Summer season Yue, a security and alignment director at Meta Superintelligence, posted on X last month describing how her OpenClaw agent ended up deleting her complete inbox, though she instructed it to verify along with her earlier than taking any motion.
Nonetheless, Meta appears bullish on the potential for agentic AI. Simply final week, Meta purchased Moltbook, a Reddit-like social media web site for OpenClaw brokers to speak with each other.
