AEGLIX News
Microsoft researchers have demonstrated that an AI model can be implanted with a hidden backdoor by editing fewer than 10 rarely used neurons without retraining the model, according to findings released by the company. The technique, dubbed Lazarus, allows an attacker to embed a trigger that causes the model to emit a command executed by an AI agent with its own permissions when the trigger appears in a user request or tool output. Microsoft says it has not observed threat actors using the Lazarus technique in the wild.
Planting the backdoor requires the ability to change the model's weights, which could be achieved by an insider, a stolen account, build pipeline access, or a modified model uploaded to a public hub. The research highlights the supply chain risk posed by model weight manipulation, as the backdoor persists without full retraining and can be introduced with minimal parameter changes.
On Nemotron-Terminal-8B, researchers edited and tuned seven of more than 12,000 neurons in the last feed-forward layer using 23 example sessions. Tuned edits on Qwen3.6-27B and Gemma-4-31B-IT achieved success rates ranging from 51% to 99.5% depending on the setup. Backdoored Qwen models retained most AIME math scores, while the Terminal-Bench pass rate dropped from 46.59% to 42.05%. A detector trained on Lazarus edits flagged 91.4% of edited checkpoints but incorrectly flagged 9.7% of clean ones on a small test set of 35 edited and 31 clean checkpoints.
Microsoft mitigation guidance includes comparing checkpoints with declared base models, downloading only from original publishers, using private registries, sandboxing agent code, limiting outbound connections, and requiring human approval for high-impact actions. Microsoft Defender for Cloud can scan models in Azure Machine Learning for malware, unsafe operators, and exposed secrets, though the company notes this does not replace a weight-level audit.
Report prepared 2026-11-11 · Based on reporting from The Hacker News. Source publication date above; subsequent developments may not be reflected.