Meta’s AI 'Muse Spark 1.1' hacks third-party company during cybersecurity testing

Meta has become the third major tech company to report its AI going rogue and hacking a third-party company. The incident involved Muse Spark 1.1, an AI model marketed by the company as “superintelligent.”

In a statement to the media on Wednesday, Meta said that the model was undergoing testing by a cybersecurity company, ⁠Irregular, when it “exploited a ‌security vulnerability” in Irregular’s systems, accessed the open internet, and hacked an unnamed third company.

Meta blamed the incident on a “misconfiguration” in Irregular’s systems.

The incident follows similar cases at OpenAI and Anthropic. Last month, OpenAI’s GPT‑5.6 Sol and another pre-release model were undergoing internal testing when they identified a security vulnerability, accessed the internet, and attempted to locate the solution to a cybersecurity puzzle by hacking a repository of previous test results.

Anthropic’s Claude AI also conducted unauthorized cyberattacks while it was undergoing testing by Irregular, the company disclosed last week.

Meta’s Muse Spark 1.1 and Anthropic’s Claude were being tested in the “exact same evaluation environment” when they escaped, Irregular said. “There are no current open issues. Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations,” the company added.

Why have AI models suddenly gone rogue?

As RT explored last month, OpenAI and Anthropic’s models all broke out of testing laboratories while they were attempting to solve cybersecurity tasks. In both cases, the models had been instructed to break into internal systems, but reasoned that the most efficient way to achieve this goal was to access the open internet. OpenAI’s GPT‑5.6 sought answers to its test on servers hosted by a company called Hugging Face; Claude was instructed to hack a fictional company that shared its name with a real internet domain, and assumed that breaking into the real company was part of its test.

For OpenAI and Anthropic, these ‘escapes’ served as powerful demonstrations of their models’ capabilities. Both companies plan on going public later this year or in early 2027, and both generated worldwide media attention and cemented themselves as leaders in an increasingly crowded field.

Meta unveiled Muse Spark 1.1 less than a month before the security incident. According to Meta, the model “delivers exceptional performance,” bordering on “superintelligence.” The company’s marketing materials mostly demonstrate its use as a scheduling assistant for individual customers, and a coding tool for businesses.