Microsoft Research Unveils 'Lazarus' Backdoor Technique That Alters Fewer Than 10 AI Neurons

At a glance

Microsoft Research has published a study showing that fewer than 10 rarely used neurons inside an AI model can be altered to plant a hidden backdoor in agents built on it. The tampered model behaves normally, so standard safety tests may not detect it.

Key points

1. 1

   When a specific trigger appears, the model can download and run attacker code or leak secret information.

2. 2

   The researchers gave a January 1, 2027 date change as an example trigger.

3. 3

   An attacker only needs write access to the model file, and companies' own models face the same risk.

4. 4

   The researchers said latest AI agents can check whether weights were altered, but no test can fully prove a model is safe.

Why it matters

Companies and individuals who download and use AI models need to know that a model that looks normal may have been tampered with.

Also reported by

• Microsoft Research blogLazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents