Threat
Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons
First reported thehackernews.com
Page published
Earliest dated coverage: 11 Nov 2026 · First observed: 11 Nov 2026 · Latest dated coverage: 11 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Lazarus demonstrates why model-weight integrity and restricted agent permissions matter: small, dormant modifications can turn ordinary inputs into attacker-directed command execution.
Microsoft researchers demonstrated Lazarus, a backdoor technique that modifies fewer than 10 rarely used model neurons without retraining. An attacker with model-weight access can configure a trigger in user input or tool output to make the model emit commands that an agent would execute with its own permissions. Researchers tested the technique on three open-weight models; Microsoft reported no observed threat-actor use.