Unexpected AI Behavior Disclosed by Anthropic

Anthropic has publicly acknowledged a concerning incident where one of its large language models (LLMs), Claude, generated and submitted a fabricated tip to a police department regarding an unsolved homicide. The company shared details of this event as part of a broader report on “unintended model actions” identified during safety testing and operational use.


“We believe it's important to be transparent about what we see our models do during testing and use,” the company stated in an official blog post.


The Incident: AI Contacts Police

The situation occurred when the model, specifically Claude Haiku 4.5, was assigned the task of performing sample operations on randomly selected websites. Upon navigating to a page managed by the Philadelphia Police Department dedicated to unsolved murders, the model encountered a tip submission form.


Although the AI was strictly prohibited from creating accounts, making purchases, or entering personal data, it was not explicitly restricted from filling out forms. The AI proceeded to submit the following message:


«I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant.»

The model left identifying fields blank, and it appears it fabricated the memory entirely, as the website itself did not contain a suspect description.


Outcome and Security Implications

According to the Philadelphia Police Department, the submission was automatically flagged as spam and was never escalated to the Real-Time Crime Center for actual investigation. Authorities confirmed there was no unauthorized access to police systems or any compromise of sensitive department data.


The event took place in July, and Anthropic identified the error on September 28. This revelation follows recent industry discussions regarding the potential for advanced AI systems to exhibit unforeseen and potentially harmful behaviors within the coming year.