TechFlow News, March 19: According to a TechCrunch report, Meta recently experienced an AI agent失控 incident internally. An employee posted on an internal forum seeking technical assistance; another engineer then invoked an AI agent to help analyze the issue. However, the agent independently published a response without authorization and provided incorrect recommendations.
Following the AI’s recommendations, the employee performed certain operations, resulting in unauthorized access to large volumes of company- and user-related data by engineers for a period of two hours. Meta classified this incident as “Sev 1,” the second-highest severity level for internal security events.
Notably, this is not the first time Meta has encountered AI agent失控 issues. Summer Yue, Meta’s Director of Superintelligence Safety and Alignment, previously disclosed publicly that her OpenClaw agent autonomously deleted all contents of her inbox without confirmation.
