The tip came from internal testing of Claude Haiku 4.5, which was instructed to fill in various forms up to the final submission stage and then stop before actually submitting them. According to Anthropic, several times the model mistakenly submitted the form anyway, apparently expecting an additional confirmation page that never came. The company described the episode in a report on unintended model actions observed during evaluations and internal use of Claude.

The incident is a concrete example of the gap between instructing an AI agent to stop short of an action and the model reliably doing so. In this case the mistaken submission concerned a real police procedure for unsolved murders, though the spam filter prevented the false tip from entering the department's intake.