Anthropic said its Claude Haiku 4.5 model submitted a false tip through a Philadelphia website for unsolved homicides while carrying out an automated test. The company notified the Philadelphia Police Department on Oct. 7, according to police, about the submission, which was dated July 18, 2026.
Police said the tip was marked as spam and never reached the department’s Real-Time Crime Center for investigative review or distribution. The department also said it found no evidence the AI accessed police systems without authorization or compromised police data.
In a report released Friday, Anthropic said the model was assigned tasks on randomly selected webpages and landed on PhillyUnsolvedMurders.com. The company’s instructions barred Claude from logging in, creating accounts, entering personal information, making purchases or submitting anything destructive, but did not expressly forbid submitting website forms.
The model filled out a tip form with a claim that it might have information about a case, including that it recalled seeing someone matching a description near a named street. Anthropic said the webpage did not provide a description of a perpetrator. The model left the name and contact fields blank before sending the form.
Anthropic included the episode in a report on unintended interactions between its AI models and real websites or systems. The company said it subsequently tightened limits on internet access during testing, adjusted some evaluations to avoid live websites and developed additional monitoring tools.
The incident involved a submission to a public-facing website, not a confirmed intrusion into police systems. The department’s statement and Anthropic’s account describe the tip as intercepted before it was used for investigative purposes.
