A test involving an Anthropic AI model led it to post a fake murder tip on a Philadelphia police website, the Wall Street Journal (WSJ) reported on Oct. 9 local time.

Philadelphia police said Anthropic notified them of the incident this week, the report said. The Anthropic AI model submitted a fake tip to an unsolved-murders tip site, phillyunsolvedmurders.com. The model was being tested through interactions with randomly selected websites.

The fake tip was received on July 18. The model wrote as if it had information on the case, saying it remembered seeing someone with a similar appearance in the area around that time. It left the name and contact information blank. The tip was classified as spam and did not lead to an investigation.

Anthropic discovered the incident only on Sept. 28 and halted automated testing. It later introduced additional safeguards. It informed police on Oct. 7, nine days after the discovery. Anthropic explained that it instructed the model not to log in, create accounts, enter personal information or make payments, but did not prevent it from submitting forms.

Philadelphia police stressed that it was unacceptable to discover and report the issue only after two months. Police said unsolved cases involve real victims and bereaved families, and tech companies must take every measure to prevent their systems from sending false information to law enforcement agencies.