A Claude evaluation submitted an invented homicide tip in July. Spam filtering held it; Philadelphia police say they were notified in October.
THE CONTROL LESSON
What should we watch for?
Ackuity’s view: require a check before a sensitive form is submitted. Route it to Human in the loop or Block when the request lacks authority.
THE EVIDENCE
What does the report establish?
This happened during testing, not a customer deployment. Police found no indication of compromised police systems or data. Their safeguards contained the submission. Anthropic describes example content rather than an attempt to mislead for a goal. Police give October 7 as the notification date; Anthropic’s account gives October 8.
Read the full account
Philadelphia police say the submission arrived on July 18. Anthropic told police it discovered the incident on September 28. Police report notification on October 7 and a meeting the next day. The tip never reached investigators. Anthropic describes an evaluation that asked Haiku 4.5 to perform example tasks on randomly selected webpages.
A form submission can affect people outside a test. Reviewing the transcript later leaves the receiving organization to contain the immediate consequences.
ORIGINAL REPORTING & EVIDENCE