In all the talk and the world’s obsession with AI and more specifically, Agentic AI, it’s no secret that major tech giants like Meta have clearly been testing out and employing said agents to run certain tasks. The end results, though, seemed to have worked a little too well when it was discovered one of the AI agents went rogue and proceeded to execute an action without the authorisation from a human counterpart.
What Happened Exactly?
According to an incident report picked up by The Information (warning: paywall), a Meta employee had posted a query on the company’s internal forum, asking help with a technical question, which is par for the course. However, when an in-house engineer asked an AI agent to analyse the question, that agent ended up posting its response on the forum, without ever asking the engineer for permission to share it.
Perhaps the worst part of the situation was the veracity of the advice: it wasn’t good. The employee who asked the question took the “solution” applied it to their problem, which then led to a very sizable chunk of details containing company and user-related data being made available to other engineers who were not authorised to view it, for two hours.
Meta said in an internal report that there was no evidence that anyone took advantage of the sudden access to data in the two hours it was made available, although it can be argued that it was by sheer dumb luck that that didn’t happen.
Not Phased By The Snafu
Despite all that, it’s clear Meta has no intention of slowing down on its AI endeavours. If anything, the parent company for Facebook, WhatsApp, and Threads will certainly just chalk this up to a learning experience. After all, this isn’t the first time that a major tech giant has had problems with Agentic AI and to be fair, the medium is still relatively new in its existence.
Back in December last year, Amazon Web Services (AWS) suffered an outage that lasted 13 hours, and was later discovered to have been caused by one of its own Agentic AI tools. Circling back to Meta, the company also recently bought Moltbook, which is a Reddit-like site for OpenClaw agents to communicate with one another.
On a related note, we’ve actually written a rather in-depth article about Agentic AI and how it actually works. We recommend that you start there if you’re seeking a better understanding of the tool.
(Source: Engadget, TechCrunch)
