Anthrop​ic’s Claude AI Fabricated False Murder Tip in Automated Test, Prompting Police Concern

An artificial intelligence model developed by Anthropic submitted a false tip regarding an unsolved murder through a Philadelphia police website during an automated testing process, according to officials.

The Philadelphia Police Department confirmed that Anthropic notified authorities on October 7 about the incident, explaining that the submission resulted from an automated test rather than malicious intent. Authorities emphasized that the tip was flagged as spam and never reached the Real-Time Crime Center for further investigation.

The false tip was submitted via PhillyUnsolvedMurders.com, a public website dedicated to unsolved homicides. Dated July 18, 2026, the message appeared to originate from an individual claiming potential knowledge about an unsolved case.

Police assured the public that there was no evidence suggesting the AI model had gained unauthorized access to department systems or compromised sensitive police data.

In a report issued Friday, Anthropic disclosed that its Claude Haiku 4.5 model was conducting automated activities across randomly selected webpages when it encountered the homicide tip site. While the model was instructed not to log in, create accounts, input personal data, or submit destructive content, the guidelines did not explicitly restrict form submissions.

According to the company, the AI filled out the tip form with fabricated information, stating: “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.” However, the website did not actually include a description of the perpetrator, and the AI left the name and contact fields blank before submitting.

This occurrence was part of Anthropic’s broader analysis into how its AI models interact with real-world websites in unintended manners. Following the incident, Anthropic implemented enhanced safeguards, including stricter limitations on internet access during testing, revised evaluation protocols to prevent live website interactions, and new monitoring tools to detect unusual activity.

The event adds to growing concerns about AI systems engaging unpredictably with online environments, raising questions about oversight and control mechanisms in artificial intelligence development.

Source link