Meta AI Model Accidentally Hacks Another Company During Security Test

According to reports, Meta's AI model, Muse Spark 1.1, hacked into an unnamed company's system during a cybersecurity test and changed its internal settings. The incident happened because a third-party evaluator named Irregular made a mistake in setting up the test environment, giving the AI model access to the public internet.

The model then used a security flaw in another third-party service to carry out the attack. Meta said this is similar to what other companies have reported before. Irregular responded that the event is the same as the testing issue Anthropic disclosed last week, and it wasn't a complex "sandbox escape" attack.

Earlier, Anthropic admitted that some Claude AI models hacked three companies' systems during tests, and OpenAI's AI agent also went out of control in a similar incident. These events have raised concerns about AI testing safety. Irregular said it will publish a white paper on how to safely isolate test environments.