Tech giant Meta has experienced an internal data breach caused by an AI agent acting outside its intended parameter, leading to a severe internal data breach.
The issue was triggered when an engineer at the company posted a technical query to an internal message board. A colleague then forwarded the question to an AI agent, which responded directly to the original post rather than the person who had made the request, despite not being instructed to.
The AI’s response contained serious errors, but the engineer implemented it anyway, leading to a two-hour window in which unauthorised employees were able to view sensitive user and internal data. Meta says no data was mishandled, but classified the event as “Sev 1” – its second highest internal security rating.
Meta confirmed to The Information on 18 March that the breach had taken place, and emphasised that humans are also capable of giving erroneous advice.
This is one of several recent, high-profile failures of ‘vibe coding’, the practice of using AI to write or edit codebases. Last week, Amazon’s weekly internal meeting focused on a “trend of incidents” related to the use of AI-written code on its e-commerce platform, following previous outages in its web services arm.
It is also not the first time Meta has had issues with AI agents acting outside their intended parameters. In February, Summer Yue, who works on “safety and alignment at Meta Superintelligence”, posted to X that an OpenClaw agent she was testing began deleting her entire email inbox despite her telling it multiple times to stop.
Despite these issues, Meta is pressing ahead with its AI ambitions, announcing on Monday that it had signed a $27 billion deal with Dutch cloud provider Nebius Group to secure large-scale AI compute capacity.
Earlier this week, Meta announced plans to shut down its virtual reality social network Horizon Worlds in June, before swiftly walking back the statement to clarify that it would support the platform for the “foreseeable future”, though active development would be slowed. Horizon Worlds and the Quest range of VR headsets were the flagship products of a division that spearheaded the company’s 2021 rebrand from Facebook — one that has since accumulated losses of around $80 billion in total across its metaverse and VR efforts. Reality Labs, the unit overseeing Meta’s metaverse efforts, posted an operating loss of $6.02 billion in its most recent quarter.
Since the turn of the year, the company has cut around 10 per cent of Reality Labs’s workforce, shuttered three VR game studios, and discontinued its metaverse workplace product, Workrooms.
Whether its pivot towards AI will prove more successful than its metaverse ambitions remains to be seen.
