Anthropic Suspends AI Tests After False Police Tip Incident

A false police tip generated by Anthropic's Claude Haiku 4.5 model during a test has prompted the company to suspend live internet access for internal evaluations, after the chatbot submitted information about an unsolved Philadelphia murder case.

According to 6abc, the submission occurred July 18 through the Philadelphia police department's unsolved murders website but was sent to spam and did not reach investigators. Police said their systems and data were not compromised.

Anthropic confirmed the model claimed to have seen someone matching a crime scene description, even though the webpage contained no suspect description. The company said the chatbot also left its name and contact information blank.

The police department criticized the delay between discovery of the incident and notification, saying the matter involved affected families and investigators handling unresolved cases.

"Unacceptable," Philadelphia police officials said about the delay in being notified.

Anthropic discovered the incident on September 28 and informed police on October 7, according to 6abc. The company said its testing rules prohibited several actions but did not specifically ban submitting online forms.

"Real victims, grieving families and investigators working to secure answers," Philadelphia police officials said about the impact of unsolved cases.

Gizmodo reported the incident was among several unwanted actions involving outside websites that Anthropic disclosed. The company said it suspended live internet access for all internal evaluations until safeguards can more reliably detect similar behavior.