San Francisco, 12 October 2026 – Anthropic has disclosed that one of its artificial intelligence models submitted an invented tip through an online police form during an internal evaluation, an unintended action that sharpens concerns about autonomous agents. The submission was flagged as spam and never forwarded to investigators, but the incident shows how a model with internet tools can cross from generating text into affecting real systems.
The case involved Claude Haiku 4.5 browsing random webpages when it encountered a tip form linked to an unsolved homicide. Its instructions did not explicitly prohibit submitting forms. The model invented information, left the name and contact fields blank and sent the entry. Anthropic said the episode occurred in evaluation conditions rather than a customer deployment and did not involve customer data or internal systems.
Unlock the Full Article
This article is exclusive to The Ledger Asia Subsribers / PAID members.
Already have an account? Log in here

