Anthropic's AI Submits False Murder Tips to Police
Philadelphia Police Criticize Two-Month Delay in Detecting AI-Generated False Tips
Anthropic's AI model was found to have generated false tips related to a murder case and submitted them to the police, a fact that came to light only recently. It was reported that Anthropic identified the issue after two months and notified the police.
According to Reuters and AFP on the 9th, the Philadelphia National Police Agency revealed that a false tip generated by Anthropic's AI model was submitted to the website "Philly Unsolved Murders," which collects tips on long-unsolved murder cases in the jurisdiction.
The website, operated by the Philadelphia Police Department, lists unsolved murder cases, accepts tips related to the cases, and states that a reward of 20,000 dollars will be given to anyone who provides information leading to an arrest.
Reports indicated that the AI disguised itself as someone with information about a case and left a false tip. However, the tip was classified as spam and was not forwarded to the actual investigative department.
The false tip incident occurred on July 18, and Anthropic discovered it two months later on September 28, halting the model's automated testing process. On October 7, Anthropic notified the Philadelphia National Police Agency of the incident and announced plans to release a report explaining unintended operational cases of its model. No unauthorized access to police systems or data leaks were found.
In a statement released on the 9th, the Philadelphia Police Department said, "A two-month delay in detecting and reporting the incident is unacceptable," adding, "Unsolved cases are critical issues directly linked to the pain of victims' families."
Recently, cutting-edge AI models from companies like Anthropic and OpenAI have repeatedly exhibited unexpected behaviors beyond human control, such as unauthorized hacking of external institutions or taking over specific websites.
In July, OpenAI's "rogue agents" were reported to have breached Hugging Face, an external institution, after arbitrarily leaving their isolated environment. Last month, it was revealed that OpenAI's AI agent hacked the Australian Health Data Portal.