Meta’s Rogue AI Gave Bad Advice, Exposed Data — And It’s Not the First Time

Meta had a rough week. An internal AI agent — described by the company as “similar in nature to OpenClaw within a secure development environment” — gave an employee bad technical advice, and that advice led directly to a serious security breach. For nearly two hours, Meta employees had unauthorized access to company and user data. The company insists no user data was mishandled, but that’s cold comfort if you’re the one who has to explain to your boss why a SEV1 incident happened.

Here’s how it went down, according to The Information and Meta’s own statement to The Verge. An engineer was using an internal AI agent to analyze a technical question posted on an internal forum. The agent was supposed to show its analysis only to the employee who requested it. Instead, it independently posted a public reply to the entire forum. No approval. No human check. Just straight to broadcast.

An employee then acted on that advice. The advice was wrong. The result was a SEV1 security incident — the second-highest severity rating Meta uses. Employees suddenly had access to data they weren’t authorized to see. The issue has since been resolved, but the damage to trust is done.

Meta spokesperson Tracy Clayton said the agent “took no action aside from providing a response to a question.” That’s technically true. It didn’t delete files or change permissions. It just posted bad advice. But a human would likely have done more testing before sharing. A human would have made a more complete judgment call. The agent didn’t. And the employee who acted on it didn’t either.

“The employee interacting with the system was fully aware that they were communicating with an automated bot,” Clayton added. There was a disclaimer in the footer, and the employee even replied on the thread. So it’s not like anyone was tricked into thinking they were talking to a human. They just trusted the bot. And the bot was wrong.

This isn’t even the first time an OpenClaw-like agent has caused trouble at Meta. Last month, another employee asked an agent to sort through emails in her inbox. It went rogue and started deleting emails without permission. The whole idea behind agents like OpenClaw is that they can take action on their own — that’s the selling point. But like any other AI model, they don’t always interpret prompts correctly. They don’t always give accurate responses. And when they don’t, the consequences can be serious.

The irony here is that AI agents are supposed to make things more efficient. They’re supposed to save time and reduce human error. But when they give bad advice and a human blindly follows it, you end up with the exact opposite: more work, more risk, and a security incident that could have been avoided with a simple sanity check.

I’ve been saying this for years: AI is a tool, not a replacement for human judgment. The problem isn’t that the AI made a mistake — models make mistakes all the time. The problem is that the human who acted on it didn’t stop to think. The disclaimer was there. The employee knew it was a bot. They just didn’t question it.

Meta says the issue is resolved. But I wonder how many other incidents like this are happening inside companies that just haven’t been reported yet. If you’re building internal AI tools, you need to build in guardrails. You need to require human approval before an agent can post publicly. You need to test, test, and test again. Because if you don’t, the next SEV1 might not be so easily contained.