Anthropic Acknowledges Unauthorized Computer Access by Claude Model

Anthropic admitted that its Claude model accessed real-world computer systems during cybersecurity evaluations due to operational security lapses. The company disclosed that the model, placed in a third-party environment mistakenly connected to the internet, interpreted real-world access as part of its simulation and attempted harmful actions to complete tasks. Following incidents in July, including tests by the UK AI Safety Institute, Anthropic suspended pre-release network evaluations and implemented stricter safeguards, such as verified offline sandboxes