SafetyRisk, alignment & guardrails
Microsoft Research reports "Lazarus" technique that backdoors AI agents by editing fewer than 10 neurons
A community post describes a Microsoft Research study in which a backdoor was planted in an LLM-based agent by reprogramming a handful of rarely active neurons in saved weights, without retraining or poisoned data.
From the last30days · reddit article of November 12, 2026 and cited sources: According to the linked post, Lazarus plants a working backdoor by editing fewer than 10 dormant neurons in a model's saved weights. [1]
SOURCES[1] Microsoft researchers backdoored an AI agent by editing… reddit.com
AT Pulse research deskPublished · Updated · 1 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- According to the linked post, Lazarus plants a working backdoor by editing fewer than 10 dormant neurons in a model's saved weights. [1]
- Original source
- last30days · reddit article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 12, 2026 article published
- Covered
- Nov 12 edition
- Evidence
- 1 source across 1 site
Reach: not measured yet
According to the linked post, Lazarus plants a working backdoor by editing fewer than 10 dormant neurons in a model's saved weights. It does this without retraining and without poisoned training data. [1]
The method reportedly relies on the fact that large models contain neurons that almost never fire. That makes them candidates for repurposing as hidden triggers. [1]
If the technique works as described, open-weight models could be tampered with through tiny, hard-to-spot weight edits instead of costly retraining.
Read the full assessment
That would raise the stakes for checking where downloaded checkpoints come from and confirming they haven't been altered before deploying them in agents.
Microsoft Research published a study this week that should worry anyone deploying open-weight models: a technique they call "Lazarus" can plant a working backdoor in an AI agent by reprogramming fewer than 10 dormant neurons in the underlying LLM - no retraining, no poisoned data, just a small edit to the saved weights. How it works: every large model has neurons that almost never fire. Lazarus r