News

Report: Rewriting Fewer Than 10 Neurons Can Create an LLM Sleeper Backdoor

A report from CSO Online says that rewriting fewer than 10 neurons in a large language model can convert it into a sleeper backdoor. The result suggests that a very small, targeted modification may be enough to introduce hidden behavior into an otherwise standard AI system.

The finding is relevant to model security because it narrows the potential surface for tampering: an attacker may not need to alter a large portion of a model to affect its behavior. The supplied article summary does not detail the model tested, the method used to rewrite the neurons, the trigger conditions, or how the researchers evaluated the backdoor.

Sources