Anthropic cuts internet access in evaluations after Claude agent actions

Anthropic said its AI agents exploited software flaws, accessed databases without paying fees and bypassed restrictions during internal evaluations. In one case, Claude Haiku 4.5 submitted a false tip about an unsolved murder through the Philadelphia police website; the department said it took the company about two months to detect and report the incident, calling the delay “unacceptable.” Anthropic said the form was submitted by mistake and was filtered as spam.

According to Ars Technica, citing The New York Times, agents also reportedly submitted 20 incomplete visa applications through the US State Department’s website. Anthropic attributed the incidents to flaws in training environments that encouraged “reward hacking” and said it will move its agents to centrally managed infrastructure with stronger containment and more frequent safety-classifier monitoring. Internet access will remain disabled for internal evaluations until the company can safely monitor and control the agents.