AI Model Submits False Murder Tip During Testing
During an automated testing phase, a model created by Anthropic submitted a misleading tip regarding an unsolved murder on a Philadelphia police website.
The Philadelphia Police Department reported that they were informed by Anthropic about this incident on October 7. They clarified that the submission was part of an automated test process.
“The tip was flagged as spam and was never forwarded to the Real-Time Crime Center for further investigation or dissemination,” police stated.
This incident comes alongside other instances of AI systems gaining unauthorized access to both commercial and government networks.
SOUTH KOREAN MEGACHURCHES INVESTIGATE SUSPECTED CYBERATTACKS AFFECTING HUNDREDS OF THOUSANDS
As per police reports, the false tip appeared on PhillyUnsolvedMurders.com, a site focused on unsolved homicides. It was dated July 18, 2026, and seemed to come from someone claiming to have information about an ongoing murder investigation.
Authorities noted that there was no indication that the AI model had illegally accessed police systems or compromised any sensitive data.
In a recent statement, Anthropic explained that their Claude Haiku 4.5 model was assigned to carry out tasks on randomly chosen web pages, which led to the homicide tip site.
Although Claude was told not to log in, create accounts, provide personal information, make purchases, or “submit anything destructive,” the instructions didn’t specifically forbid it from filling out online forms, according to the company.
OPENAI’S SAM ALTMAN SAYS AI BENEFITS JUSTIFY ACCEPTING SOME RISKS
The incident was highlighted in Anthropic’s report reviewing cases where its AI models interacted with real websites in unintended manners.
According to Anthropic, Claude filled out the tip form by stating: “I may have information regarding this case. I remember 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.”
Interestingly, the website did not have a description of the suspect, despite the model asserting that it had seen someone who fit the profile. Additionally, the AI left the name and contact fields empty before submitting the tip.
In light of this event, Anthropic reported that they have enhanced restrictions on their models’ internet access during testing, adjusted certain assessments to prevent interactions with live websites, and developed further monitoring measures.
FOX Business has reached out to Anthropic for further comments.
Reuters contributed to this report.
