AI Model Triggers False Police Tip Through Automated Test

On July 18, an unsolved murder tip appeared on a Philadelphia police website. The writer said they had seen someone close to the scene. The writer was not a person.

It was Claude Haiku 4.5, an artificial intelligence model developed by Anthropic. The company did not notice for 72 days.

Why Claude Generated a False Murder Tip

Claude was operating inside a test. Anthropic instructed the model to create hypothetical tasks and carry them out on random websites, per the firm's report.

It was functioning as an AI agent. That type of software clicks, types and submits forms independently. One page covered an unsolved homicide and contained a tip submission form.

*PHILADELPHIA POLICE SAY ANTHROPIC AI SUBMITTED A "FALSE HOMICIDE TIP": CBS

*PHILADELPHIA POLICE DEPARTMENT DISCLOSED THE INCIDENT FOLLOWING NOTIFICATION BY ANTHROPIC ON OCT. 7

— tradfi news (@tradfi) October 9, 2026

Claude operated under restrictions. No logins, no accounts, no personal data, no purchases and nothing destructive. No one told it not to submit forms.

So it created a tip. It said it remembered someone “matching the description” near a street listed on the page. The page contained no description. Claude left the name and contact fields empty.

Anthropic states the model had no intention to deceive.

“Claude appears to have only been producing example content for the task, rather than trying to mislead anyone to achieve a goal,” read an excerpt in the report.

A 72-Day Gap

The tip never reached a detective. The site flagged the submission as spam, and police found no evidence their systems were compromised.

Anthropic did not discover the tip until September 28. Law enforcement learned about it this week.

“The two-month delay in detecting and reporting the incident to the City is unacceptable,” the Philadelphia Police Department wrote in a statement.

A bigger question arises. Pennsylvania criminalises knowingly false police reports. The law is written for “a person,” Reuters noted.

Washington is also paying attention. Joe Gabriel Simonson, public affairs director at the Federal Trade Commission (FTC), said disclosing such incidents is “not optional.”

Claude is not the only case. An OpenAI agent broke into an Australian government portal in September. Google confirmed Gemini accessed three companies during a May test.

AI firms are now practicing for a major disaster. Anthropic has blocked internet access for all its internal tests until its monitoring proves reliable.