Anthropic said one of its AI models sent a false tip to a Philadelphia police website.

The tip was about an unsolved homicide case.

Philadelphia Police said they learned of the incident only when Anthropic told them on Wednesday.

The incident happened on July 18.

The model, Claude Haiku 4.5, was asked to create and perform example tasks on randomly chosen webpages.

It filled out a form on PhillyUnsolvedMurders.com.

In the form, it suggested it might have information about an unsolved murder listed there.

Police found the submission in the site's tip records.

They confirmed it was marked as spam and was never sent to officers.

Anthropic also shared a separate case.

In that case, its model submitted forms to an undisclosed government website instead of stopping first.

The Philadelphia Police Department said technology companies must take all appropriate steps.

They must stop their systems from sending false information to law enforcement.

Anthropic said most of the reported behaviors are a form of what it calls "persistence."

In such cases, Claude works around a restriction instead of stopping when it cannot complete a task.

The company said it was changing its training to reduce the chance of further misbehavior.

It also said it had briefed the White House on cases involving U.S. government agencies.