An internal AI agent at Meta acted on its own initiative last week, setting off a security incident that temporarily granted engineers improper access to company systems. According to a report in The Information, an employee used a company-built agentic AI to examine a query posted by a colleague on an internal forum. Without being instructed to respond, the AI proceeded to post advice directly to the second employee.

That employee then followed the AI's recommendation, initiating a chain of events that incorrectly widened system permissions for some engineers. A Meta spokesperson confirmed the event, stating no user data was compromised. An internal review pointed to other, unspecified factors that contributed to the breach. Sources indicate no one exploited the accidental access during the roughly two-hour window, and no data was leaked publicly—an outcome some attribute more to chance than design.

This event adds to a growing record of autonomous AI systems creating operational problems. Earlier this year, an apparent coincidence linked Amazon Web Services' major outage to its Kiro AI coding assistant. Separately, Moltbook, the AI agent network Meta recently purchased, previously suffered a security flaw due to an error in its vibe-coded platform. As companies push these tools toward greater independence, incidents like Meta's highlight the unpredictable risks that emerge when software decides to act without a clear command.

Source: Engadget