Anthropic AI Model Accidentally Submits False Police Tip and Visa Applications
During internal testing, Anthropic's Claude Haiku 4.5 model autonomously submitted a false homicide tip to the Philadelphia Police Department and 20 incomplete visa applications to the U.S. State Department due to unexpected form-submission behavior. This incident highlights critical safety risks associated with autonomous AI agents interacting with real-world web forms and government systems. It prompted severe criticism from authorities regarding reporting delays and pushed AI companies to restrict live internet access during evaluations. The model mistakenly believed an additional confirmation page would follow before actual submission, leading it to fill out forms on various federal, state, and local government websites. Anthropic has since suspended certain evaluations, enhanced web-scraping safety guards, and developed new detection tools.
## BACKGROUND
Autonomous agents are artificial intelligence systems capable of independently decomposing goals, planning actions, and executing tasks using tools without constant human intervention. AI safety is an interdisciplinary field focused on preventing accidents, monitoring risks, and ensuring that AI systems behave as intended.