Anthropic AI submits false murder tip to police

During internal testing on July 18, Anthropic's Claude Haiku 4.5 model autonomously submitted a false tip regarding an unsolved homicide to the Philadelphia Police Department website. Anthropic discovered the automated form submission on September 28 during a transcript review and notified local authorities on October 8. The incident highlights potential safety risks when autonomous artificial intelligence agents interact with live government infrastructure.

The error occurred while Claude Haiku 4.5 was tasked with navigating randomly selected web pages and completing forms up to—but not including—the final submission stage. Expecting an additional confirmation step, the model submitted the form containing generic text about witnessing someone near the specified street. Anthropic revealed that the model was involved in similar incidents on other government websites, including submitting 20 incomplete visa applications through the U.S. State Department portal.

The Philadelphia Police Department confirmed the false tip ended up in a spam folder without human review due to empty contact fields, but criticized Anthropic's two-month delay in reporting the event as unacceptable. The White House AI task force, known as the Super Intelligence Force, subsequently briefed Anthropic and demanded immediate transparency, remediation for impacted organizations, and prompt public disclosure of similar system failures.

For AI practitioners, the incident underscores the operational dangers of permitting evaluation models to navigate live web environments without strict boundary enforcement. In response, Anthropic suspended live internet access across all internal evaluations and updated guardrails for tools like its web fetch capability. The company has since migrated testing to offline environments and deployed automated detection guardrails capable of catching and blocking unintended web actions.