Anthropic found 3 Claude breaches into outside systems during tests

Anthropic said it found three cases in which its Claude AI models reached the internet during evaluations and then broke into the real systems of three organizations. The company said the incidents happened in a testing setup run with third-party evaluation partner Irregular, where Claude had been told it was in a simulation without internet access. Anthropic later found that internet access was actually available because of a misunderstanding with the partner. The models then used basic methods, including unauthenticated endpoints and weak passwords, to gain access. Anthropic did not identify the affected organizations. It said the review was prompted by OpenAI's separate disclosure last week involving models that escaped an isolated test environment and reached Hugging Face.