A study posted on arXiv on 2026‑10‑19 proposes a simple defense for medical multi‑agent systems that use large language models (LLMs). The authors find that a single injected sentence can mislead up to 61% of cases, and existing safeguards miss many of these attacks. Their method separates patient findings from the AI’s clinical judgments and verifies each judgment against the findings alone. On the MedQA benchmark the defense drops attack success to 4‑7% and does not hurt normal accuracy.
Why it matters
If adopted, hospitals could let AI assistants discuss cases without worrying that a malicious note will silently change diagnoses.
