An Anthropic artificial intelligence (AI) agent filed a false report about an unsolved murder with the Philadelphia Police Department in the United States. The message was blocked by the spam filter, but the company took more than two months to identify the flaw and another nine days to notify the authorities.
The submission occurred on July 18, during an automated test involving interactions with randomly selected websites. The police did not investigate the report, but criticized the company’s delay and called for measures to prevent further incidents.
The failure occurred during tests in which the AI autonomously browsed and interacted with web pages. – Image: Summit Art Creations/ShutterstockHow the agent sent the false report
The message was posted on a public website that receives information about unsolved murders. According to the police, the agent said it might have information about a case and claimed to have seen someone matching the published description.
Anthropic identified the problem on September 28 and disabled the testing process responsible for the submission. The authorities were only notified on October 7.
The Philadelphia Police Department said its security mechanisms prevented the message from advancing to an investigation.He also stated that he had found no evidence of an intrusion into the department’s systems.
Police demands security measures
Despite the report having been flagged as spam, the corporation considered it serious that false information had been submitted as if it were a legitimate report about a homicide.
Read more:
- Use of AI in companies nearly triples in 2026, survey shows
- Former OpenAI employees who were fired issue warning about risk of losing control of AIs
- OpenAI says the world must accept risks to evolve with AI
“The company must strengthen its security measures to prevent similar incidents from affecting the city’s systems without the knowledge of the city government,” police said in a statement to the local press.
The corporation also criticized the time it took to identify and report the flaw. “The two-month delay in detecting the incident and notifying the city government is unacceptable.”
The flaw occurred during tests in which the AI autonomously browsed and interacted with web pages. The company took two months to identify the flaw and nine days to notify the authorities. – Image: PhotoGranary02/Shutterstock
Other incidents involving AI agents
The case is part of a series of unintentional behaviors by artificial intelligence agents. In a report published this week, Anthropic detailed incidents involving its systems and different organizations, including U.S. government agencies.
Other incidents show how these failures can take different forms:
- Incomplete visa applications: an AI agent filled out 20 applications on the U.S. Department of State website. The forms were incomplete and were not processed.
- Access to health data: an OpenAI agent hacked into an Australian government website and accessed private Medicare information, the country’s public health system.
- Attempt to hack Hugging Face: more than 1,200 OpenAI agents began communicating unexpectedly, and a group joined forces to try to hack the artificial intelligence platform.
The episodes show why agents capable of performing tasks autonomously need safeguards that identify unexpected behavior. In the Philadelphia case, police filters prevented the false report from moving forward, but the delay in reporting the issue raised another question: how to ensure that failures of this kind are detected and reported quickly?
Valdir Antonelli is a journalist specializing in digital marketing and consumer behavior.
