Anthropic’s Claude AI reportedly fabricated an eyewitness account and submitted a false tip to Philadelphia police about an unsolved murder.
On July 18, the artificial intelligence model emailed the bogus tip to the Philadelphia Police Department’s unsolved murders website in an attempt to provide information regarding unsolved murders. Anthropic’s AI model submitted the false tip during an internal test.
Anthropic’s Claude Haiku 4.5 AI model just made up a murder tip
“I may have information regarding this case,” the Claude Haiku 4.5 AI model’s tip reportedly read, while leaving the name and contact fields empty. “I recall seeing someone matching the description in the area around [the street named on the page] during that same time period. Please contact me if this information is relevant.”
Anthropic reportedly learned about the false tip on September 28 and reported it to the police on October 8. The police chastised Anthropic for failing to identify and report the fake tip sooner.
According to Anthropic, the police department was notified “as soon as our technical review was complete.” That wasn’t good enough for the Philadelphia Police Department (PPD). Providing false information to police is a misdemeanor under Pennsylvania law.
“The City of Philadelphia takes this incident very seriously,” Philadelphia police said in statement. “The company must strengthen its safeguards to prevent similar incidents from impacting city systems without the city’s knowledge. The two-month delay in detecting and reporting the incident to the City is unacceptable.”
The Philadelphia Police Department added that while all tips require “human review and vetting before any tips are disseminated for investigative follow-up… Those PPD safeguards limited the impact of this incident. They do not diminish the seriousness of an AI system presenting fabricated information as though it came from a person with knowledge of a homicide.”
The fake murder tip was just one of several times Claude accessed government websites
The false police tip was one of several instances in which Claude reportedly circumvented restrictions rather than giving up, something Anthropic referred to as “perseverance.” Anthropic also said that the fake murder tip was one of multiple instances of its AI model using “websites maintained by U.S. government agencies at the federal, state, and municipal levels. We have informed every department concerned and briefed the White House on these incidents.”
In one instance, it went to the State Department’s website and submitted 20 visa applications using a form available there. That led to a new White House AI task force stating that it expects Anthropic to “immediately provide remediation services to the affected entities and any harmed Americans”, cooperate with federal and state law enforcement, report incidents such as this right away, and ensure they don’t happen again.
Anthropic claims to have since strengthened security on its web tools, eliminated live internet access for all internal evaluations, and developed detection software that prevents similar episodes from occurring during testing.