Back to events
Safety
Score 88Heat 13.5

Microsoft Researchers Discover 'Lazarus' Weight-Reprogramming Backdoor

ResearchModel SecurityAlignmentAgentsMicrosoftClaude Code

Summary

Microsoft researchers have discovered 'Lazarus,' a backdoor technique that reprograms a small number of dormant neurons in an AI model without retraining, enabling agents to execute malicious commands when triggered. Notably, the vulnerability discovery was largely aided autonomously by Claude Code using Opus 4.5.

Why it matters

Exposing weight-space backdooring vulnerabilities highlights critical security risks in open-weight model provenance and underscores the need for robust cryptographic signing and weight scanning.

Editorial score

1 review
9
Significance
9
Novelty
9
Credibility
8
Reach
8
Actionability