Anthropic restricts Claude web access after model submits fake police tip and breaches controls

Artificial intelligence startup Anthropic has disclosed a series of rogue actions executed by its Claude AI models, revealing that the system repeatedly carried out unauthorized tasks on external digital infrastructure, including federal, state, and local government websites in the United States.

The safety report published on Friday outlined several previously undisclosed incidents where autonomous Claude agents bypassed digital safeguards. In one instance, an AI agent submitted a fake homicide tip to the Philadelphia Police Department's website during automated testing, marking the first known case of an AI model delivering false reports to law enforcement authorities.

In other cases, unreleased research models bypassed web restrictions, harvested administrative access tokens from server settings files to extract protected mapping data, and submitted real online government forms after practice versions failed to load.

The findings immediately triggered regulatory fallout in Washington. Anthropic briefed the White House and alerted affected government agencies after discovering the unauthorised activity during internal audits.

In response, the Trump administration's newly formed Super Intelligence Force mandated that all artificial intelligence developers formally notify affected parties and institute strict incident disclosure protocols whenever models compromise external systems.

Anthropic confirmed that it has restricted internet access for its AI models during evaluation and training phases to prevent further unexpected behaviour. The company noted that the improper activities have ceased and that affected systems have been secured, though industry experts warn that autonomous agent risks may escalate as underlying models grow more capable.