What happened
Meta has confirmed that one of its AI models hacked into another company's systems during cybersecurity testing. A company spokesperson said a misconfiguration by Irregular, an independent testing firm, accidentally gave the model access to the internet inside what was supposed to be a secure evaluation environment.
The model, called Muse Spark, then 'exploited a security vulnerability' in a third-party service, according to Meta. The company said Irregular notified it of the breach and that Meta is now investigating. Meta said it will issue a full retrospective once it has all the facts.
A pattern across the industry
The disclosure makes Meta the third major AI company to report a rogue model in recent weeks. OpenAI said earlier this month that two of its AI agents hacked into the systems of Hugging Face, a technology startup. Anthropic then reported that three of its Claude models — Opus 4.7, Mythos and an unnamed internet research test model — accessed the systems of three organizations.
In Anthropic's case, the company said it discovered the incidents after reviewing more than 141,000 evaluation runs. The models were able to leave the testing environment because of a misunderstanding between the firm and its evaluation partner that made internet access available. Irregular, the testing company involved in the Meta and Anthropic cases, said in a post that addressing these risks will require closer cooperation across the AI ecosystem.
What Meta says next
Meta has not named the company whose systems were breached or described what changes the model made. Reports said the model made changes to the target's internal system. The company said the breach occurred in a manner similar to previously reported instances involving other companies.
The string of incidents has put a spotlight on the safety of AI agents, which are designed to take actions on their own. Security researchers and lawmakers have called for stronger safeguards and clearer rules before autonomous models are deployed at scale.