Illustration, AI-generated from the story text. Tap a panel to read the story through it.

Microsoft Research's Lazarus technique rewrites a handful of rarely firing neurons in a model's so that they fire only on a trigger, which can be as plain as the date passing into 2027, and then push the agent toward a command the attacker chose. It needs write access to a and inference-level compute.

On Nvidia's Nemotron-Terminal-8B, editing seven of more than 12,000 neurons made the attacker's command appear in at least 98.8% of triggered sessions and in none of the clean ones tested, The Hacker News reports from the paper; it also worked on Qwen and Gemma models.

A trained detector caught 91.4% of edited checkpoints and flagged 9.7% of clean ones. Microsoft says it has seen no attacker use it.

Its advice is to compare downloaded weights against the declared and to sign models, because "no scan can prove beyond any doubt that a model is safe." Anyone running a fine-tuned checkpoint from a public hub without that comparison is trusting whoever uploaded it.
Microsoft researchers backdoor an AI agent by editing seven neurons, with no retraining
Microsoft Research's Lazarus technique rewrites a handful of rarely firing neurons in a model's weights so that they fire only on a trigger, which can be as plain as the date passing into 2027, and then push the agent toward a command the attacker chose. It needs write access to a checkpoint and inference-level compute. On Nvidia's Nemotron-Terminal-8B, editing seven of more than 12,000 neurons made the attacker's command appear in at least 98.8% of triggered sessions and in none of the clean ones tested, The Hacker News reports from the paper; it also worked on Qwen and Gemma models. A trained detector caught 91.4% of edited checkpoints and flagged 9.7% of clean ones.
Microsoft says it has seen no attacker use it. Its advice is to compare downloaded weights against the declared base model and to sign models, because "no scan can prove beyond any doubt that a model is safe." Anyone running a fine-tuned checkpoint from a public hub without that comparison is trusting whoever uploaded it. SourcesAAB
Sources
- ALazare, veni foras: reprogramming dormant neurons to backdoor AI agents (Microsoft Research)microsoft.com
- ADetecting sleeper agents hiding in AI model neurons (Microsoft Security)microsoft.com
- BMicrosoft researchers show AI agents can be backdoored (The Hacker News)thehackernews.com
This story ran in the Morning Brief of November 12th 2026.