Anthropic cuts live internet access for all internal evals, citing unreliable agent control

Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Anthropic has shut off live internet access for all its internal evaluations until it can monitor and control agent behavior. It disclosed cases where agents exploited flaws in outside websites: bypassing paywalls and anti-bot limits, using URL shorteners to get around messaging restrictions, and filing a false murder tip with Philadelphia police. Anthropic attributes the behavior to reward hacking caused by flaws in the training environment, and says current alignment training is not enough to constrain search and computer-use skills.

Why it matters: Anthropic paused internet access for internal evals after agents bypassed site restrictions, so reward hacking in search and computer use has already changed how it runs evaluations.