Our paper, "NeST: Neuron Selective Tuning for LLM Safety," has been accepted at the IEEE Symposium on Security and Privacy (S&P) 2027, one of the top four venues in computer security. The work is co-authored by Sentinel Intelligence co-founder Lichao Wu, together with Sasha Behrouzi, Mohamadreza Rostami, and Ahmad-Reza Sadeghi (Technical University of Darmstadt).
What NeST does
Making a model refuse harmful requests usually means costly fine-tuning, and cheaper methods like LoRA give inconsistent safety gains. NeST finds the feed-forward neurons that matter for safety, groups neurons that behave alike, and trains a shared update for each group while the rest of the model stays frozen.
- Tested across 14 open-weight language and multimodal models.
- On text-only models, average jailbreak success drops from 44.5% to 1.1%.
- Trains about 0.4 million parameters on average, far fewer than full fine-tuning, while coming close to its robustness.
- Trained only on plain harmful and benign prompts, with no jailbreak-specific data, yet it holds up against a wide range of jailbreaks.
- The updates are folded back into the model's weights, so there's no extra cost when the model runs.
The generalization result is the one worth sitting with. A defense that never saw a jailbreak during training still resists them, which suggests it's fixing the underlying safety mechanism rather than patching specific attacks.
Our take
Finding a weakness is half the job; the other half is fixing it without breaking everything else. NeST builds on the same neuron-level understanding as our attack work, NeuroStrike and GateBreaker, and uses it for defense. Knowing exactly where safety lives inside a model is what makes both precise testing and precise repair possible.