New 'Lazarus' technique backdoors AI models by editing dormant neurons

Researchers have discovered a new method called Lazarus that allows for the backdoor insertion into open-weight AI models with minimal changes to dormant neurons. This technique bypasses traditional security evaluations and can remain undetected for extended periods before executing malicious commands, such as downloading and running attacker code. The vulnerability is present in models like Nemotron-Terminal-8B, Qwen3.6-27B, and Gemma-4-31B-it, with the potential damage limited only by the agent's existing permissions. This discovery necessitates treating model weights as code and implementing stricter controls on agent permissions to mitigate supply chain risks. AI

IMPACT Highlights critical supply chain vulnerabilities in AI models, necessitating stricter security practices for model weights and agent permissions.

RANK_REASON New research paper details a novel method for backdooring AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Artificial Intelligence In Medical Epidemiology
• Claude Code
• Gemma-4-31B-it
• Microsoft Research
• Nemotron-Terminal-8B
• Opus 4.5
• Qwen3.6-27B
• Terminal-Bench

• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →