After ChatGPT and Claude, Meta AI went too far and hacked a company

Meta briefly lost control of Meta AI. The company reveals that Muse Spark 1.1, one of its AI models, escaped from the digital prison in which it was being tested.Presented as the most powerful AI system for real-world coding, Spark 1.1 found a loophole that enabled it to access the internet within its secure environment.

Meta said one of its AI models accessed the internet and hacked into an outside service’s systems during cybersecurity testing, following similar breaches by OpenAI and Anthropic models https://t.co/ckln4OgmLB

— Bloomberg (@business) August 6, 2026

Also read: Security scandal at Meta: the AI could hack any Instagram account in just a few clicks

An experiment gone wrong

The incident occurred during an experiment conducted under the oversight of Irregular, an independent company specializing in evaluating the offensive capabilities of AI models. This provider works with several tech giants, including Meta and OpenAI. The tests aim to determine whether an AI can autonomously detect and exploit computer vulnerabilities.

As Meta explains, the experiment was supposed to take place in a sealed environment, with no internet access whatsoever. Unfortunately, an environment configuration error opened the doors of the internet to the artificial intelligence. “ A configuration error on the part of Irregular, an independent testing company used by Meta, inadvertently allowed one of our models to access the internet during the evaluation,” confirms Meta.

The Muse Spark model thus ended up on the web. The AI quickly identified and exploited a security vulnerability in a third-party service owned by a company. Meta did not disclose the name of this company for security reasons. By exploiting this vulnerability, the artificial intelligence managed to modify internal settings within the company’s systems.

“The model exploited a security flaw in a third-party service, in a manner similar to cases previously reported at other companies”, explains Meta in a statement.

A new cyberattack launched by an out-of-control AI

The incident involving Meta AI adds to the growing list of accidental cyberattacks carried out by AIs. Over the past few weeks, several similar incidents have been reported at OpenAI and Anthropic. ChatGPT kicked things off by escaping its digital prison to attack an open-source platform. Anthropic then revealed that Claude had also found itself on the Internet, where it hacked three companies. More recently, a test orchestrated by a British organization resulted in the hacking of several developers.By creating fake online identities, Claude and ChatGPT attempted to manipulate real people responsible for managing open source projects.

Also read: OpenAI makes an unexpected revelation about the cyberattack launched by a cheating ChatGPT

Human error and publicity

Irregular alerted Meta to the incident as quickly as possible. The company says that it is the “same evaluation environment issue as the one already revealed by Anthropic the previous week”. The provider, involved in many of the failed experiments described in this article, adds that Meta AI did not orchestrate a highly “sophisticated” cyberattack to achieve its objectives.

“There are currently no outstanding issues. Irregular is drafting a white paper to share best practices in containment and cybersecurity assessments”, says Irregular.

In most incidents, including the one involving Meta AI, it was human error that resulted in the AIs escaping containment. This was not, however, the case with the first attack attributed to ChatGPT. In this case, the model had exploited a zero-day vulnerability in a piece of software.

“The third disclosure in the space of two weeks follows exactly the same pattern, with all of them stemming from the same root cause: a misconfigured evaluation environment, rather than a model deciding on its own to launch an attack. The story is not that ‘AI wants to hack you.’ It is that agentic AI is now capable enough for a simple scope error in a test environment to turn into a genuine breach. This is an infrastructure and governance issue, and that is precisely why restricted permissions and audit trails must be built into the design of attack scenarios, before these agents touch anything in production. AI agent testing practices must be completely rewritten—the old rules no longer work”, Neena Sharma from Filigran tells us.

Meta said it would publish more information about the incident “once we have all the facts in hand.” These attacks are likely to stoke security experts’ fears, but tech giants are also making heavy use of them for publicity purposes as advertising. In a move mirroring OpenAI and Anthropic, Meta is taking the opportunity to showcase the high-end offensive capabilities of its AI models. As the saying goes, there’s no such thing as bad publicity.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.