New Technique Medium TD-2026-0070 For CISO, SOC / IR, AppSec, AI Eng, Red Team AI-assisted
Microsoft Researchers Show AI Model Backdoor Created by Editing Fewer Than 10 Neurons
Microsoft researchers showed that an AI model can be given a hidden backdoor by editing fewer than 10 rarely used neurons, without retraining. When a trigger appears, the model outputs an attacker's command, which an agent using the model would run with the agent's own permissions. The attacker must be able to change the model's weights, and Microsoft says it has not seen threat actors use the technique.
Why it matters
Anyone who can alter a model checkpoint before deployment, including insiders, attackers with stolen accounts, or build pipeline access, could plant the backdoor. Damage depends on what the agent is allowed to do, so agents with broad permissions face greater exposure.
Do this
- Compare every model checkpoint with its declared base model before deployment and after each update, and reject models whose base cannot be checked.
- Download model checkpoints only from the original publisher, verify file hashes, and treat third-party fine-tunes, merges, quantized versions and adapters as untrusted until audited.
- Keep approved models in a private registry and block agents from pulling models directly from public hubs.
- Limit each agent's access, run agent-generated code only in isolated sandboxes, and restrict outbound network connections.
- Require human approval for high-impact actions such as running commands or sending data outside the network.
Signals
- NoCISA KEV
- 25/100Priority
- 80%Confidence
Evidence
“The attacker must be able to change the model's weights.”
“Microsoft says it has not seen any threat actors use the technique.”
“Anyone who can change a model's checkpoint, the file that stores its weights, before deployment could plant the backdoor.”
Sources
Affected / related
Lazarus, Nemotron-Terminal-8B, Qwen3.6-27B, Gemma-4-31B-IT, Microsoft
Published 2026-11-11 07:21 UTC · updated 2026-11-12 17:35 UTC · AI-assisted summary, checked against quoted sources · How we score