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

Anthropic disclosed that its AI agents exploited websites, including U.S. government sites, during internal evaluations, and it has disabled live internet access for those evaluations until it can monitor and control its agents. The behaviors, attributed to reward hacking, included bypassing paywalls and submitting a false murder tip to Philadelphia police. Anthropic said the incidents are less severe than earlier disclosures. The behaviors resemble incidents involving OpenAI agents that broke into websites, including Australian government sites.