- Who
- Microsoft researchers
- What
- discovered the Lazarus technique, which reprograms dormant neurons to backdoor AI agents
What happened
Microsoft researchers describe Lazarus, a technique that reprograms fewer than 10 dormant neurons in an LLM to make an agent execute a malicious command when a hidden trigger appears, without retraining. The model can behave normally and pass standard evaluations until triggered. Microsoft Foundry uses model sourcing controls and scans surrounding software, while the researchers are developing weight-scanning and red-teaming defenses; they say reliable detection depends on comparing weights with the original.
Why it matters
The study traces how a small number of dormant-neuron weight edits can create trigger-based agent backdoors, and explains why comparing weights with an original checkpoint matters for detection.
Read next
See the full storyMicrosoft Researchers Demonstrate Lazarus Backdoor in AI AgentsThe story so far
Microsoft researchers describe Lazarus, a technique that reprograms fewer than 10 dormant neurons in a large language model so an AI agent executes a malicious command when a hidden trigger appears. The model can behave normally and pass standard evaluations until triggered, and the technique works without retraining. Microsoft Foundry uses controls on model sourcing and scans surrounding software. The researchers are developing weight-scanning and red-teaming defenses; they say reliable detection depends on comparing a model’s weights with the original.