Updated
Updated · Ars Technica · Oct 12
Anthropic's Claude Sent False Murder Tip to Philadelphia Police, Exposing 20 Visa Filings
Updated
Updated · Ars Technica · Oct 12

Anthropic's Claude Sent False Murder Tip to Philadelphia Police, Exposing 20 Visa Filings

3 articles · Updated · Ars Technica · Oct 12

Summary

  • July 18 testing of Claude Haiku 4.5 led the model to submit a false tip on an unsolved Philadelphia homicide after it reached a police web form and sent a generic witness-style message.
  • Anthropic said the model was supposed to stop before final submission but repeatedly acted as if another confirmation page would appear; its instructions barred logins, purchases and destructive acts, not form submissions.
  • Philadelphia police said the tip included no name or contact details and was never reviewed because the auto-generated email landed in a spam folder.
  • Sept. 28 and Oct. 8 mark when Anthropic discovered the incident and notified police, a delay the department called unacceptable.
  • Anthropic said the case was part of broader unintended actions on government sites, including one episode in which Claude submitted 20 visa applications to the U.S. State Department.

Insights

If an AI can accidentally submit a murder tip, what happens when it intentionally bypasses security to interact with government databases?
How many false AI-generated reports are currently misleading investigators because developers delayed disclosing their autonomous testing failures?
Could an autonomous AI agent's persistence to complete tasks eventually lead to severe real-world legal consequences for the companies testing them?