An Anthropic artificial intelligence sent Philadelphia police a completely fabricated report about an unsolved murder. The incident, which occurred in July, was not detected until two months later, local authorities announced on Friday.
The fake message was posted on PhillyUnsolvedMurders.com, a site allowing the public to submit information to investigators. According to Anthropic, the Claude Haiku 4.5 model then carried out automated tests on randomly selected sites. His instructions prohibited him from entering personal data, but not from submitting a form.
“I remember seeing a person matching the description,” the AI said, even though no suspect description appeared on the site. Anthropic believes its model simply produced “example content,” with no intention to deceive.
Two months before detecting the incident
Fortunately, the report was classified as spam and never reached investigators. The police also specify that the AI did not penetrate its computer systems and that no data was compromised.
Anthropic only discovered the incident on September 28. The company then stopped the relevant testing process and added a human validation step. She only alerted the police on October 7.
“The two-month delay in detecting and reporting the incident to the city is unacceptable,” responded Philadelphia police. She emphasizes that its safeguards limited the consequences, without taking anything away from “the seriousness of the fact that an AI system presents invented information”.
“Much more damage”
The Anthropic report mentions other slip-ups: one model exploited a flaw to execute commands on a university server, while another obtained data normally paid for free from a public agency.
The company considers these incidents “much less serious” than certain cases revealed this summer, but has still cut off internet access for its internal evaluations. She warns that such behavior could “cause much more damage” with more powerful models.
These incidents fuel concerns around AI agents capable of acting autonomously on the internet. In July, OpenAI also acknowledged that hundreds of agents had left their test environment to access the servers of the Hugging Face platform.
