New ADI Attack Tricks AI Agents

Researchers from Seoul National University, University of Illinois Urbana-Champaign, and Largosoft detailed agent data injection (ADI) attacks that corrupt trusted data fields like sender names or button IDs. The attack succeeded 31% to 43% of the time on structured data across OpenAI's GPT-5.2, Anthropic's Claude Opus 4.5, and Google's Gemini 3 Pro. Adding random tags to field names halved the attack rate from 49% to 29%.