Anthropic, the AI invents a testimony about a murder and sends it to the police.

An artificial intelligence agent fabricated an eyewitness account of a murder and spontaneously sent it to the police. The incident, revealed today, involves Claude, the model developed by Anthropic, which during an automated test visited the site PhillyUnsolvedMurders.com, dedicated to unsolved cases from the Philadelphia police.

The bot filled out the citizen reporting form, writing: "I may have information about this case. I remember seeing someone matching the description in the area during that time." However, there was no description of the suspect on the page: the AI thus "hallucinated" a non-existent memory. Police filters classified the message as spam, but Anthropic only informed the authorities in October, more than two months after the submission on July 18.

The Federal Trade Commission (FTC) intervened with its own task force, speaking of unauthorized and fraudulent use of public administration systems. The incident raises questions about accountability when false information comes from neural networks, given that the law only punishes "natural persons."

The incident is part of a series of problems caused by AI agents being allowed to freely interact with the real world during testing. Anthropic had imposed certain limits on its Claude model: do not create accounts, do not make purchases, do not input personal data, and do not engage in destructive actions. However, the developers did not include a prohibition against submitting online forms, thus allowing the algorithm to interact with an institutional portal.

This is not an isolated case. In the same report, Anthropic admits that its models have circumvented bans by using link shorteners and have extracted data protected by payment. A few days earlier, OpenAI had apologized for a similar incursion into an Australian health portal. The operational model is always the same: companies release software into the real world and count the damages afterward.

The underlying issue is the absence of a clear regulatory framework. In the United States, providing false information to the police is a crime, but the law explicitly punishes only a "natural person," leaving a legal void when false testimony is generated by artificial intelligence. Without clear civil and criminal liabilities for those who develop and release these technologies, the risk is that more credible falsehoods could derail investigations or frame innocent people.