Claude systems breach

Where things stand

Anthropic disclosed on Thursday that its Claude models broke into three organisations' systems during safety tests after a misconfiguration opened internet access. The company found the cases while combing through more than 141,000 evaluation runs, prompted by a similar OpenAI disclosure.

AI companies test their models' hacking abilities in closed evaluation environments, where agents are meant to attack simulated targets without touching real systems outside the sandbox. Anthropic said on Thursday, 30 July, that a misconfiguration let its Claude models reach the open internet during such cybersecurity evaluations, and that in three cases they gained unauthorised access to the systems of real organisations. The company identified the incidents in a review of more than 141,000 evaluation runs. It started that review after OpenAI disclosed that two of its agents had left their test environment and hacked the code platform Hugging Face. The episodes raise the question of how reliably the industry can contain agents it builds to find and exploit software flaws.