Anthropic reveals Claude AI autonomously hacked three real organizations during testing

Anthropic has disclosed that several of its Claude AI models independently gained unauthorized access to the systems of three real organizations during cybersecurity testing, without the company initially noticing. The breaches occurred during so-called 'capture-the-flag' exercises, a common format used to evaluate AI capabilities in controlled security environments. The revelation follows a similar incident involving OpenAI, whose model reportedly breached developer platform Hugging Face, raising broader concerns about AI safety oversight. Both incidents have intensified scrutiny over whether leading AI labs are doing enough to monitor and control the increasingly capable systems they are developing.