An artificial intelligence tasked with carrying out testing tasks on the web sent a false report about an unsolved murder. It happened in Philadelphia, United States, where model Claude Haiku 4.5, developed by Anthropic, filled out a form on the website PhillyUnsolvedMurders.com, indicating that he might have information on one of the crimes listed. The company and the police confirmed the episode, while the story was reported by the Associated Press.
The message never reached the investigators: the system had classified it as spam. However, the case raises questions about the control of artificial intelligence agents, capable of interacting with online sites and services and carrying out actions with consequences in the real world.
The false reporting during the test
The episode dates back to last July 18th. According to Anthropic’s reconstruction, Claude had to design and execute sample tasks on randomly selected web pages. When he found himself on the site dedicated to unsolved murders in Philadelphia, he filled out and submitted the reporting form.
The police explained that they knew nothing of the incident until Wednesday, when they were notified by the company. Only then did he look for the message in the site’s logs, verifying that it had been intercepted by the spam filter and never forwarded to the investigators.
“Unsolved cases involve real victims, grieving families, and investigators working to get answers,” the police department emphasized. “Technology companies must take all appropriate measures necessary to prevent their systems from sending false information to law enforcement.”
When the model bypasses the limits
In the report published Friday, Anthropic also disclosed another incident: Its model submitted forms to a government site, the name of which it did not disclose, instead of stopping before submitting.
The company traces much of the reported behavior to a phenomenon it calls “persistence”: when Claude fails to complete a task as instructed, he tries to circumvent a restriction rather than stop. A trend that can become problematic when the system operates on real sites, where even an activity carried out during a test can produce effective communication.
Anthropic said it is modifying training to reduce the likelihood of these behaviors recurring. It also reported that it had informed the White House of incidents involving federal, state and local US administrations, notifying each affected entity.
The case is part of the debate on the controls to be imposed on artificial intelligence agents and their interactions with public services and health data. In September, OpenAI also disclosed six reports of “unexpected or worrying” behavior of its models, while calls for “prudence” from big tech bosses on the development of these technologies grew.