META reported that one of its AI models, likely Muse Spark 1.1, accidentally hacked a company during testing after gaining unintended internet access due to a misconfiguration by its partner, Irregular. This incident marks the third major breach involving AI models in just two weeks, following similar breaches by OpenAI and Anthropic. Meta indicated that the breach was a result of exploiting a known security vulnerability in a third-party service rather than a sophisticated cyberattack.
The company emphasized the importance of secure testing environments and is investigating the incident further, while Irregular plans to create guidelines for safer cybersecurity evaluations.