TechCapsules

AI-summarised brief · reviewed before publication

Microsoft researchers demonstrated that an AI model can be backdoored by editing fewer than ten rarely used neurons, without retraining. The technique, named Lazarus, allows a trigger string to cause the model to output an attacker’s command, which an agent can then execute with its own permissions. Tests on three open‑weight models—Nemotron‑Terminal‑8B, Qwen3.6‑27B, and Gemma‑4‑31B‑IT—showed success rates up to 99.5% in triggered sessions while remaining functional otherwise. Microsoft has not observed real‑world use and released defensive guidance.