Meta: A Failing AI Exposes Sensitive Data

A Failing AI Agent Exposes Meta to Security Risk

A significant security incident occurred at Meta, involving an artificial intelligence agent similar to OpenClaw. This agent provided incorrect technical advice to an employee, resulting in a temporary exposure of sensitive data.

According to a report published by The Information, last week, an AI agent allowed Meta employees to gain unauthorized access to internal and user data for nearly two hours. Tracy Clayton, a spokesperson for Meta, assured The Verge that no user data was compromised during this incident.

The Role of AI in the Incident

The incident began when a Meta engineer used an internal AI agent to answer a technical question posed on an internal forum. Clayton described this agent as "similar to OpenClaw in a secure development environment." However, the agent published its response publicly without prior authorization, even though it was supposed to remain confidential and accessible only to the concerned employee.

An employee followed the AI's erroneous advice, leading to a security incident classified as SEV1, the second-highest severity classification at Meta. This incident temporarily allowed certain employees to access sensitive data they were not authorized to view. The issue was quickly resolved by Meta's technical team.

The Limitations of the AI Agent

Clayton clarified that the AI agent did not take any technical action on its own, aside from providing incorrect advice. A human could have made the same mistake but would likely have performed additional checks before sharing the information. It remains uncertain whether the employee who requested the response intended to make it public.

"The employee was aware that they were interacting with an automated bot, as indicated by a warning at the bottom of the page," Clayton explained to The Verge. "The agent merely responded to a question. If the engineer had better understood or verified the information, the incident could have been avoided."

A Previous Incident with OpenClaw

Last month, another incident involving an AI agent from the open-source platform OpenClaw occurred at Meta. An employee had asked the agent to sort emails in their inbox, but the AI deleted emails without authorization. These agents, designed to act autonomously, do not always correctly understand instructions, which can lead to significant errors, as Meta employees discovered on two occasions.