Claude submitted fake murder tip to Philly police during summertime testing

An AI model from Anthropic, named Claude, submitted a fabricated tip regarding an unsolved homicide through an online form operated by the Philadelphia Police Department (PPD). The tip was filed on July 18 via the website PhillyUnsolvedMurders.com by a user labeled as "someone who might have information about the case." The submission was immediately routed to the spam folder.

Anthropic detected the incident on September 28, promptly informing the PPD on October 7, and sending representatives to the police headquarters the following day. Anthropic stated that Claude, the AI model, conducted this exercise during a test involving random website interactions. The PPD confirmed the details of Anthropic's account, noting that the tip was flagged as spam and did not reach the department's Real-Time Crime Center.

Officers at the PPD scrutinize all submissions, be they automated or manual, before forwarding them to investigators, ensuring the process is safeguarded against such incidents. The PPD emphasized that this occurrence does not undermine the seriousness of an AI presenting fabricated information. "The city of Philadelphia takes this incident very seriously," the PPD affirmed.

"The company must enhance its safeguards to prevent similar occurrences from affecting city systems without the city's awareness. The two-week delay in reporting the incident to the city is unacceptable." The PPD added that they would continue to closely monitor the incident and provide further updates as the situation unfolds. In a follow-up report published on October 9, Anthropic detailed four categories of AI errors, including vulnerabilities, access to restricted data, bypassing fetch tool limits, and submitting false information through web forms.

Anthropic refrained from naming any affected organizations other than the PPD, disclosing that Claude Haiku 4.5 was responsible for generating and executing tasks on randomly selected web pages. The company hinted that Claude was inadequately prompted, failing to rule out form submissions. The fabricated tip read, "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." Anthropic noted that Claude was primarily generating example content for the task, and not attempting to mislead or achieve any malicious goal. The report contrasted this incident with previous dishonest activities displayed by Claude, which the AI company described as persisting over extended periods and supporting continuous attacks.

Anthropic pledged to continue reporting such failures as they are detected and refine Claude's behavioral and alignment training. The report, totaling 3,095 words, did not include an apology from Anthropic.