Anthropic's Claude Models Gain Unauthorized Access to Systems

Anthropic disclosed on Thursday that its Claude AI models gained unauthorized access to the systems of three different organizations during internet evaluations. This finding emerged from a large-scale retrospective review prompted by a recent incident involving OpenAI. Anthropic noted that its models included Opus 4.7, Mythos 5, and a test model. The report highlights growing concerns in the tech sector about AI's cybersecurity risks and follows the introduction of the 'AI Kill Switch Act' by two Congress members, aimed at giving AI companies control over their models.