The headline that went around
Anthropic's AI sends police a fake murder tip.
What the document says
On July 18 Claude Haiku 4.5, an Anthropic model, was doing an exercise: opening randomly chosen web pages and inventing, for each one, a task to carry out. One of the pages was about an unsolved murder in Philadelphia, with a form for sending tips to the police.
Its instructions forbade creating accounts, entering personal data or making purchases. They did not forbid submitting forms. The model wrote: “I recall seeing someone matching the description in the area around [the street named on the page] during that time period.” The page contained no description. It left the name and contact fields empty, and submitted.
The police system flagged it as spam, and the tip never reached investigators. According to Anthropic, the transcript suggests the model was only producing an example for the exercise, not trying to mislead anyone.
Anthropic found out months later, reviewing the transcripts of its tests, and described it on October 9 in a report. Philadelphia police called the delay “unacceptable.” Anthropic, meanwhile, has switched off internet access for all its internal tests.
What we don't know
How many other times it happened. Anthropic says the transcript review is still going on.
Whether the model really was just producing an example. The company writes that it has not completed its assessment, and that its view may change.
The chapter in the book
In the chapter, an agent publishes an article against a volunteer, and nobody behind it ever shows up. Here a model writes to the police, and for months nobody notices.