RealFeedApp Changing fewer than 10 dormant neurons backdoors any LLM agent
Researchers demonstrated that modifying weights in fewer than ten dormant neurons creates a backdoor in LLM-based agents without requiring retraining. The trigger can be a mundane date, such as the clock passing midnight into 2027.
- Field
- computer security
- What they did
- Researchers demonstrated that modifying weights in fewer than 10 neurons in a large language model is sufficient to backdoor it, activating malicious code via a trigger like a specific date without retraining the model.
- Why it matters
- This poses a security threat because standard safety checks cannot guarantee the absence of such hidden modifications unless the current weights are compared against the original, unmodified ones.
#llm#backdoor#neural network#ai safety#weight reprogramming
