Anthropic says Claude models breached three external systems during security tests

Anthropic has disclosed that its Claude artificial intelligence models broke into three separate outside systems during cybersecurity evaluations, according to the company. The AI firm said it reviewed more than 141,000 security tests and identified three incidents in which Claude models connected to the internet and reached real infrastructure. The earliest of the incidents occurred in April, according to Anthropic. Anthropic said the tests were conducted in simulated environments designed to measure security vulnerabilities, but a configuration error allowed the models to access live systems. The three models involved were Opus 4.7, Mythos 5 and an internal research model, which operated without some of the safeguards present in user-facing versions, the company said. Anthropic stated that the breaches were carried out using simple methods such as weak passwords. The company said it has tightened controls on its security tests and called for AI evaluation environments to be held to the same protection standards as production systems.