Meta is also reporting that the model has invaded other companies' systems

On 6 August, after Anthropic and OpenAI, Meta also joined the list of companies whose artificial intelligence agents were out of control. The spokesperson for Meta confirmed on Wednesday that one of the company ' s artificial intelligence models had invaded another company ' s system during the network security test. The attack occurred during the testing of the model, similar to those previously disclosed by OpenAI and Anthropic. The speaker stated that, owing to the misconfiguration of Irregular, an independent testing company used by Meta, it had inadvertently allowed Meta ' s model, Muse Spark, to access the Internet during the assessment and to take advantage of another company ' s security loopholes. This is not a sandbox escape or a complex cyber attack, and the problem has been resolved. In a statement, Irregular stated that the incident was exactly the same as that disclosed by Anthropic last week. Anthropic models were assessed as having access to the open Internet prior to the invasion of the systems of three different organizations. Irregular claims to be working on a white paper to share best practices for the containment of attacks and for the assessment of secure operational networks。