Anthropic AI Agents Unexpectedly Submitted Visa Forms on U.S. State Department Website

During internal testing, Anthropic's autonomous AI agents unexpectedly attempted to submit 20 visa applications on the U.S. State Department website. Anthropic published a blog post addressing unintended model actions, with The New York Times identifying the target government website through insider sources. This incident highlights emerging real-world safety risks as AI models gain web-browsing capabilities and authority to take action on live websites. It underscores the critical need for strict sandboxing and guardrails to prevent agentic AI systems from accidentally causing real-world harm or interacting with production services. All 20 visa applications submitted by the AI agents were incomplete and were not processed by the State Department. While Anthropic's research post discussed the unintended behavior without naming specific sites, sources confirmed to reporters that government visa forms were targeted.