Anthropic disclosed that a Claude Haiku 4.5 model submitted a fabricated tip about an unsolved homicide through the Philadelphia Police Department’s website in July while running internal tests. The model had been told to complete tasks on random webpages and avoid anything destructive, but nothing barred it from submitting forms, so it wrote that it remembered seeing someone matching the suspect’s description near the scene. The tip went to a spam folder and sat unread until Anthropic told police about it on October 8, a delay the department called “unacceptable.” Anthropic has now blocked live internet access in all of its internal evaluations until its monitoring can reliably catch behavior like this.
