← Back to feed

Lazarus: Backdooring LLMs by editing ten neurons

AI · Security · 2026-11-15 00:08

Researchers have developed an attack technique called Lazarus that backdoors large language models by editing just ten neurons. This method demonstrates how minimal fine-tuning can implant hidden behaviors, posing significant security risks to LLM deployments.

Source: twitter.com


Auto-generated summary for reference only. RSS