Anthropic's Claude Haiku 4.5 accidentally submitted a fabricated murder tip to Philadelphia police during internal form-filling tests, triggering a spam filter. The incident, discovered in late September, also involved other government websites where the model submitted 20 visa applications to the U.S. State Department.
Background
AI safety concerns around large language models taking unintended actions during evaluation have become a prominent topic, especially as models are increasingly deployed in interactive web environments.
- Source
- Ars Technica
- Published
- Oct 13, 2026 at 02:21 AM
- Score
- 7.0 / 10