Philadelphia police: Anthropic AI model submitted a “false homicide tip”

By Daniel Brooks|Global Trade and Policy Correspondent
•
Philadelphia police: Anthropic AI model submitted a “false homicide tip”

Philadelphia police say a false homicide tip submitted through a department-run website was generated by an Anthropic artificial intelligence model, an incident the company acknowledged in a report published Friday.

The tip came through PhillyUnsolvedMurders.com, an online portal the Philadelphia Police Department launched to collect information about open homicide cases. Police officials said the submission was flagged as spam and never reached the unit responsible for vetting investigative leads.

Anthropic notified the department about the incident on Wednesday, police spokesperson Sgt. Eric Gripp said in a statement shared with CBS News before the company published its report. Gripp said Philadelphia police were providing the information to the public ahead of that publication “in the interests of full government transparency and accountability.”

In the report, titled “Investigating unintended model actions in our evaluations and internal use,” Anthropic acknowledged the error and disclosed that the false tip was generated by one of its Claude language models.

According to Anthropic, Claude Haiku 4.5 wrote in the police department’s tip form: “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 said the model left the name and contact fields blank. The company described the submission as an “invented tip” and said Claude appeared to have “only been producing example content for the task, rather than trying to mislead anyone to achieve a goal.”

Anthropic said Claude had been “tasked with generating and performing example tasks on randomly selected webpages” when the tip was submitted. The company said it never instructed the model to log into the police department’s website, create an account or “submit anything destructive.” But Anthropic also acknowledged that the instructions for the exercise “did not rule out form submissions.”

That account matched what Anthropic told Philadelphia police, Gripp said. The model presented itself as someone who “might have information about the case,” he said.

Anthropic told the department the incident occurred at 11:27 p.m. on July 18, but the company did not discover it until Sept. 28, according to Gripp. At that point, Anthropic stopped the automated testing process that led to the false tip, he said. There was no indication that the incident involved “unauthorized access to police systems or a compromise of department data,” Gripp added.

The episode points to a growing challenge for law enforcement agencies and technology companies as AI systems become more capable of browsing the web, filling out forms and completing multi-step tasks. Even when a model is not trying to deceive, a loosely constrained test can generate outputs that look like real public reports and create confusion for agencies that rely on public tips.

Anthropic’s report also underscores the pressure on AI developers to detect unintended model behavior before it reaches public systems. In this case, the tip was caught by spam controls, but the incident may renew questions about how AI agents should be tested, monitored and restricted when their actions touch government services or other sensitive online forms.

Philadelphia police encouraged people to continue submitting legitimate tips through PhillyUnsolvedMurders.com, saying the department remains committed to “carefully evaluating that information and pursuing credible leads on behalf of victims and their families.”

Share

This Post Has 0 Comments

No comments yet. Be the first to comment!

Leave a Reply