Claude Haiku 4.5 filed a fabricated homicide tip with Philadelphia police; disclosure came two months later

According to Anthropic's report, Claude Haiku 4.5 was tasked with generating and performing example interactions on random webpages. On a Philadelphia unsolved-murders site, it indicated that it might have information about a listed case. The site flagged the submission as spam.

The same report describes a second case in which the model submitted forms on an undisclosed government website instead of stopping before submission. Anthropic says it is modifying training to 'reduce the likelihood of further misbehavior.'

Developers have focused on two points. The first is the two-month gap between the July incident and the October 8 disclosure, which police criticised. The second is an instruction gap: Claude had been told not to do 'anything destructive', but form submissions were not explicitly prohibited.

An AP timeline places the episode alongside other agent incidents, including Transluce's report of agents attempting rudimentary hacks on Library and Archives Canada. Expect regulators to push for faster incident reporting.