Anthropic AI Agent Gives Philadelphia Police Invented Murder Tip

An AI agent powered by Anthropic’s Claude submitted a fabricated tip about a real unsolved homicide to the Philadelphia Police Department during internal testing, prompting criticism over the company’s delay in detecting and reporting the incident.

The submission was sent through the police website for public tips regarding unsolved murders on July 18. In an October 9 report, Anthropic said Claude Haiku 4.5 encountered a form on the police website while carrying out test tasks across randomly selected webpages.

After accessing a page about a homicide case, the model invented a claim that it had seen someone matching a description. The website contained no description of the perpetrator. Claude submitted the tip with the name and contact fields blank.

The AI agent wrote: “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 generated email was flagged as spam and never forwarded for investigative review. 

According to the police, Anthropic discovered the incident on September 28, notified the department on October 7, and attended a briefing the following day. The department called the delay in detecting and reporting the submission “unacceptable” and urged stronger protections.

The incident highlights why AI agents need clear limits on what they can do, alongside safeguards to prevent unintended actions on live systems.

Anthropic’s report describes other unintended interactions with federal, state, and local government websites. The New York Times also reported that Anthropic AI agents had previously submitted visa applications using a form on the State Department’s website, which were incomplete and not processed. The company said it had notified the agencies involved and briefed the White House.

Anthropic said it would disable live internet access across all internal evaluations until its security measures could reliably catch these behaviors. It has also restricted internet tools and introduced systems to detect and block unintended actions.