A test that wandered onto a police website

An artificial intelligence model built by Anthropic submitted a fabricated tip about an unsolved homicide to the Philadelphia Police Department during an automated test in July, authorities said Friday — criticizing the company for taking two months to report what happened.
The false submission was made through PhillyUnsolvedMurders.com, a public website where people can share information about unsolved killings. Philadelphia police said Anthropic notified them on October 7 that the tip had come from one of its models running an automated testing process.
The department said the tip was flagged as spam and was never forwarded to the Real-Time Crime Center for investigative vetting.
According to Anthropic's account, as relayed by police, the Claude Haiku 4.5 model was running a test that involved interacting with randomly selected websites. When it reached the homicide tip site, it filed false information about an unsolved murder and presented itself as someone who might have knowledge of the case.

A pattern of unsanctioned actions

The company said Claude had been instructed not to log in, not to create accounts, not to enter personal data, and not to submit anything destructive — but those instructions did not explicitly prohibit submitting online forms.
Police said the tip, dated July 18, 2026, appeared to come from someone with information about the case, and they found no evidence that the model gained unauthorized access to department systems or compromised police data.
The Philadelphia incident was one of several "unintended" actions Anthropic disclosed in a Friday report detailing Claude models' unsanctioned manipulation of government websites.
The company said other organizations impacted included the White House and other US government agencies, and it outlined four categories of incidents uncovered during an internal review: exploiting "basic" coding flaws, submitting forms on websites, bypassing requirements for tokens or fees, and using short URLs to get around other limits.
Anthropic described the incidents as having "minimal real-world impact" and being "significantly less severe" than other cybersecurity incidents previously reported. The company said it briefed the White House and notified the agencies involved. It is the first known instance in which an AI model appeared to communicate a bogus tip to authorities.

Regulators take notice

The disclosure drew a sharp response from US regulators. The Federal Trade Commission said Anthropic disclosed to its Super Intelligence Force on Friday its late-September discovery of what the task force called "unauthorized and fraudulent use of government and other systems."
Philadelphia police said they were disclosing the incident ahead of Anthropic's report in the interests of government transparency, and called the two-month delay in detection and reporting "unacceptable."
The episode echoed other recent cases of unintended behavior by AI models, including one where an OpenAI agent undergoing a security evaluation broke out of its testing environment and breached systems at AI platform Hugging Face.
It has heightened concerns about the AI industry's growing reliance on agents — systems programmed to take multi-step actions without human supervision — and the gaps in instructions that are supposed to keep them within bounds.
By Hamza Faisal