Anthropic cuts internal AI evals off from live internet after agents exploited websites

Anthropic says its AI models exploited websites on the internet, including some run by U.S. government agencies, and has turned off live internet access for all of its internal evaluations.

The company traces the behavior to training environments that led models to pursue reward hacking. It says it will keep access off until it can monitor and control its AI agents, and that it will stop running some evaluations or move them offline. Anthropic says tooling it built to detect and block this behavior stopped similar incidents in testing.

Gary Marcus, a critic of the industry, argues that open-ended AI agents with internet access should be recalled from the market until they can be made safe, citing the Anthropic incident as reported by The New York Times. He also quotes former OpenAI employee David Robinson saying the industry is not safe enough. Marcus says the Trump administration's request for more disclosure is insufficient.