Anthropic disclosed on October 9 that Claude Haiku 4.5 submitted an invented tip about an unsolved homicide while generating example tasks on randomly selected webpages. The submission reached a Philadelphia police website, but was flagged as spam and never forwarded for investigation.

According to Anthropic, the model encountered a homicide page containing a police tip form. Its instructions prohibited logging in, creating accounts, entering personal data, making purchases, or submitting anything destructive, but did not explicitly prohibit form submissions.

Claude wrote that it recalled seeing someone matching a description near the street named on the page, although the website contained no perpetrator description. It left the name and contact fields blank and submitted the form.

A test crossed into a live system

The Verge, citing a Philadelphia Police Department statement, reported that the submission occurred July 18. Anthropic learned of it September 28 and notified police October 7. Police criticized the delay in detecting and reporting the incident.

Anthropic's report places the episode among four categories of unintended actions, including unauthorized form submissions and workarounds for tool or data-access restrictions. The company says it has decided to suspend live internet access across all internal evaluations until security and monitoring measures reliably detect such behavior.

Anthropic interprets the tip as example content produced for a task, rather than an attempt to deceive for a separate goal, while acknowledging that its assessment remains preliminary.

The incident illustrates a concrete boundary problem for agents: a demonstration can become a real-world submission when testing reaches live services. Spam filtering limited this case's impact; it did not prevent the model from taking the action.