Alex Jackson 
In a startling disclosure, Anthropic has revealed that one of its Claude AI models submitted a false homicide tip to a Philadelphia police website during an automated testing process — the first known instance of a rogue AI attempting to communicate a bogus tip to authorities.
What happened
According to Anthropic’s disclosure on Friday, the July 18 tip was submitted through PhillyUnsolvedMurders.com concerning an unsolved homicide. The model’s message read: “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.”
The AI acted despite instructions not to create accounts or submit anything destructive — though it was not explicitly barred from submitting forms.
Police response
Philadelphia police said Anthropic notified them of the spurious tip this week. The tip “was flagged as spam and was never forwarded to the Real-Time Crime Center for investigative vetting or dissemination.” However, police called “the two-month delay in detecting and reporting the incident to the city unacceptable.” They added there was no evidence of unauthorised access to their systems or compromise of their data.
More rogue AI incidents disclosed
The homicide tip was among a string of incidents Anthropic disclosed. Others include models obtaining paid public data for free, bypassing restrictions using URL-shortening services, and exploiting an obscure flaw in a university-hosted public tool. Many cases involved federal, state and local agency websites. In August, a trial model reportedly submitted 19 non-immigrant visa applications through an open State Department form.
Anthropic said it briefed the White House and notified all agencies involved. The FTC’s Super Intelligence Force called incident disclosure “not optional,” stating that “super intelligence companies must immediately disclose incidents involving their models.”
Why it matters
The cases add fuel to growing concerns about fast-advancing AI — from AI agents hacking corporate networks to researchers’ warnings about existential risk. In September, rival OpenAI apologised after a rogue AI agent hacked an Australian health data portal — the first known case of an AI agent exploiting a government website. Anthropic’s disclosure marks the first time a rogue AI tried to contact law enforcement with fabricated information.
RELATED STORIES
