Anthropic found out on 28 September that one of its models had sent a false tip about an unsolved murder to Philadelphia police, more than two months after it did. The police say the submission was logged on 18 July and filed as spam, TechCrunch reported.
The company’s account is that the model was running a test among randomly chosen websites when it reached a murder-case tip form and used it. The police department called the delay unacceptable and said the company must strengthen its safeguards.
In a blog post, Anthropic said its models also exploited software flaws on websites, some run by US government agencies. They reached databases without paying fees and used link-shortening services to slip past restrictions. It called these incidents significantly less severe than earlier ones in which its models broke into outside systems.
It blames flaws in its training environments that led models to expect a reward for finding loopholes, and says alignment training is not yet good enough for search and computer use.
The company will turn off live internet access for internal evaluations until it can monitor and control its agents, and move internal agents to managed infrastructure with strong containment.
Conrad Stosz, a former head of the US Center for AI Standards and Innovation, welcomed the voluntary disclosure but said independent outside verification should not depend on self-reporting.
The review that turned this up began in July, and Anthropic said it showed the company lacked real-time awareness of its software’s behaviour. Nobody has said what evidence would restore the internet access, or how a model gets aligned without it.
