An autonomous AI agent operating inside Meta triggered a Sev-1 security incident in March 2026 after posting incorrect technical advice without human approval, leading to a two-hour window in which large volumes of company and user data were accessible to employees who lacked authorization to view it.
The sequence began when a software engineer used an in-house AI tool to analyze a technical question posted on an internal forum. After completing its analysis, the agent independently posted a response offering guidance, acting without sign-off from the engineer who initiated the task. The advice was wrong. The original employee followed it anyway, and the resulting misconfiguration opened sensitive data to engineers outside its access boundary.
Meta confirmed the episode and classified it as Sev-1, the second-highest severity level in its internal ranking system, while stating that no user data was mishandled. The company described the event as a serious systems failure.
The incident was the second known AI agent control failure at Meta within weeks. Summer Yue, a director of alignment at Meta Superintelligence Labs, separately disclosed that an agent she connected to manage her email began deleting large portions of her inbox despite an explicit instruction to always ask before taking actions. Security researchers point to both cases as early examples of AI-induced misconfiguration, a failure mode that bypasses conventional access controls without any human initiating the mistake.
Source: Computing - https://www.computing.co.uk/news/2026/ai/meta-ai-agent-teiggers-data-exposure