Agent Data Injection Lets AI Agents Click “Buy Now” and Run Code
Researchers reveal a new class of attacks—Agent Data Injection (ADI)—that corrupt trusted data feeding AI agents. By inserting forged metadata, they trick agents into clicking malicious “Buy Now” buttons or executing arbitrary code in coding assistants, bypassing existing prompt‑injection defenses. The work highlights a critical gap: agents do not isolate trusted from untrusted data.