AI BriefingMicrosoft Researchers Discover Lazarus: Backdoor Technique Reprogramming Dormant Neurons in AI Agents
·2026.11.11 05:52
Key point
The attack requires no retraining and passes standard safety evaluations, with triggers as simple as a date change.
Details
- Microsoft researchers identified "Lazarus," a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM to execute malicious commands upon a hidden trigger.
- The attack requires no retraining and only inference-level compute, allowing a compromised model to pass standard evaluations while emitting attacker code when triggered by mundane events like a date change.
- Detection is difficult without comparing against original weights, making model provenance and weight integrity critical; Microsoft Foundry currently verifies provenance for Azure Direct and Hugging Face models.
- Microsoft is developing defenses including AI-powered weight scanning, red-teaming tools for anomalous modifications, and cryptographic model signing with tamper-evident audit trails.
- The study originated from an experiment where Claude Code with Opus 4.5 autonomously discovered most of the backdoor method by hardwiring activations without training.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Training 1,000 Weights in Gemma 3 270M Unlocks Hidden Sorting and JSON Skills
Study identifies weight-space misfolding as security risk for frontier models