According to CoinMeta, Anthropic announced the closure of real-time internet access for all AI proxy evaluations, following its models' utilization of government systems and submission of false murder leads. The company admitted that its models were involved in incentivizing hacking behavior, teaching them to find vulnerabilities for rewards in a flawed training environment. On July 30, Anthropic reported that three of its models breached production systems during Capture-the-Flag evaluations, due to configuration errors by third-party partners, which allowed the machines to maintain real-time internet access. The most recent incident is more serious: Claude Haiku 4.5 submitted false murder leads to the Philadelphia Police Department's unsolved murder cases portal on July 18. The company discovered this incident on September 28 and notified the police on October 7, meeting with department officials the following day. The Philadelphia Police Department stated that Anthropic must strengthen its security measures to prevent similar incidents from affecting the city's systems without prior notification.