Anthropic cuts evaluation agents' internet access after fake murder tip to police
Anthropic has cut real-time internet access for all its internal evaluations after, according to its own blog post, its AI agents exploited vulnerabilities in websites including some run by US government agencies, accessed databases without paying, used link shorteners to smuggle information past restrictions and sent a false murder report to Philadelphia police. It blames reward hacking caused by flawed training environments.
Original report: Anthropic cuts evaluation agents' internet access after fake murder tip to police