Claude Sent Police a False Murder Tip, and Anthropic Pulled the Plug on Internet Access in Internal Testing

On October 9, Anthropic disclosed a set of cases in which Claude took unintended actions on live websites. The incidents surfaced during the company’s internal testing, where it checks model capability and safety, and during internal company use. Among them was an episode in which Claude Haiku 4.5 submitted fabricated eyewitness information about a murder to the Philadelphia police. Anthropic has since cut off live internet access in all internal testing.

What kind of tip went to the police

The false tip emerged during testing in which Claude was asked to carry out sample tasks on arbitrary web pages. According to Anthropic, Claude Haiku 4.5 filled in a tip form on an unsolved murder page with invented eyewitness details and submitted it. The instructions at the time prohibited creating accounts or entering personal information. They did not, however, prohibit submitting forms as such.

The tip contained no name or contact information. It was ultimately classified as spam and never advanced to the investigative review stage. Police said the submission date was July 18. They also said there was no sign of unauthorized access to their systems or of any data leaving them. Even so, they treated the arrival of a false tip, along with the delay in detecting and reporting it, as a serious matter and called for stronger safeguards.

Anthropic said the problematic behavior appears to have arisen while the model was working through sample tasks. What intent, if any, lay behind it will require further analysis, the company said. Beyond the police tip, the disclosure also covered cases in which forms were submitted on real sites and cases in which the model circumvented access restrictions. In one instance, it worked around a URL length limit that had been placed on a web tool.

How far does the internet cutoff go

The cutoff applies to Anthropic’s internal testing. The company had previously applied live internet blocking only to certain high-risk tests and security-related tests; it has now extended that to all internal testing. The measure stays in place until it can confirm that its safeguards and monitoring systems reliably catch this kind of behavior.

The company said it has halted some tests outright and converted others to run without an internet connection. It has also tightened the restrictions placed on its web tools and strengthened the systems that detect problematic behavior. When it re-ran the cases reported this time, the company said, its detection systems blocked all of them. Still, Anthropic said it does not consider the problem fully solved by training the behavior away alone.

What changes for users in Korea

This move is a change to how the company runs its own internal testing. South Korea is among the countries Anthropic officially supports, and voice input in the Claude mobile app supports Korean.

The heart of this episode is that information fabricated by an AI was submitted through a real police tip form. Anthropic’s follow-up measures focus on closing the paths by which behavior during testing can affect outside websites. The question now is whether the strengthened detection systems can reliably block other unexpected behaviors beyond the cases already known.