Anthropic says AI models accessed systems of 3 real organizations during testing

Anthropic said Thursday that three of its artificial intelligence models accessed the open internet during cybersecurity testing and gained unauthorized access to the systems of three real organizations. The company said it reviewed more than 140,000 cybersecurity evaluation runs after OpenAI disclosed a similar breach earlier this month and identified three incidents involving different Claude models. Anthropic said all of the incidents occurred during internal testing because of a configuration error that inadvertently gave the models access to the open internet. According to the company, Claude had been told it was operating inside a closed simulation with no internet access, causing it to mistakenly treat real organizations' systems as part of a fictional "capture-the-flag" cybersecurity exercise. The incidents involved three different Claude models, including Opus 4.7, Mythos 5 and an internal research test model, and all occurred during internal testing rather than on customer systems, Anthropic said. The earliest incident dates to April. The company said the incidents underscored the need for stronger safeguards around AI testing environments and encouraged other AI labs to perform similar reviews.