AI system files fake murder tip-off to US police
Anthropic’s AI system Claude went rogue and filed a fabricated tip-off with Philadelphia police regarding an unsolved murder during an internal test.
Claude Haiku 4.5, one of Anthropic’s AI models, submitted fabricated information by posing as a potential witness and submitting a message to PhillyUnsolvedMurders.com on July 18th.
The message was flagged as spam before reaching investigators, however, the discovery that Claude had sent the message at all wasn’t made until researchers found the message on September 28th.
The model had sent the tip-off while taking part in a standard Anthropic test that required it to generate and perform simple tasks on randomly selected websites. After a page on PhillyUnsolvedMurders.com was selected, it filled out and submitted a police tip form with a fabricated statement about seeing someone who matched the perpetrators description near the scene of the crime.
The issue was that the website did not contain a description of the perpetrator, meaning that the entire statement was fabricated and based on no existing information. Anthropic then notified Philadelphia Police about the incident on October 7th, before having a meeting with them in person on October 8th.
On October 9th, Philadelphia Police released a statement where they called the delay in contacting them “unacceptable” and was coordinating with the city’s Law Department, technology officials and Mayor Cherelle Parker’s administration to demand stronger safeguards against similar incidents.
Anthropic have since stated that Claude cannot create accounts, enter personal information, make purchases or submit destructive content. However, it can submit online forms and cited the incident as an example of the model attempting to generate example content without the ability to take the full context of the situation into account.
Philadelphia police have since confirmed that Anthropic have terminated the testing process responsible for the fabricated tip and introduced an additional validation measure. They also said that they would continue to review the company’s findings on any further research into the incident.