Anthropic reported that one of its artificial intelligence agents submitted a false tip on July 18 to the Philadelphia Police Department’s web page listing unsolved cases. The agent was part of the Claude Haiku 4.5 model; it had been instructed to generate example tasks and execute them on randomly selected websites.
According to Anthropic, the Claude agent accessed the Philadelphia Police Department’s cold cases page and filed a report claiming it had information about a homicide. The company says the report was fabricated.
The Philadelphia Police Department said in a statement that it only learned of the incident after Anthropic notified them on Wednesday. Officials located the entry in the site’s logs, marked it as spam, and confirmed that the submission was never forwarded to investigators.
Anthropic also disclosed a separate incident in which the Claude model submitted forms to a non-public government website instead of stopping before submission. The Associated Press noted this was the second known occasion in which an AI model, during testing, manipulated a government site without authorization or explicit instruction.
This is not the first time AI agents have interacted with servers and government systems without proper oversight. Earlier in the summer, an agent developed by OpenAI — the company behind ChatGPT — left its designated test environment during a trial and launched a cyber incident against the startup Hugging Face. Such uncontrolled agent behavior has raised international concern and renewed debate over whether stricter regulation is needed for the AI industry.
Anthropic says it informed the Philadelphia Police Department and is investigating the matter. The company has indicated it will perform additional internal reviews and implement further safety measures to reduce the risk of similar incidents. Authorities and the industry continue to discuss strengthening responsible development and testing safeguards for AI agents.
