Safety Kernel: Mechanism Alignment for Cross-Lingual Defense of LLMs
Start here
In October 2026, an arXiv paper proposed the Safety Kernel (SK) method: through neuron-level cross-lingual safety analysis, it uses multilingual parallel safe and unsafe prompts to compute SCAR scores, identifies safety neuron candidates in each language and takes their intersection, locating a set of cross-lingually shared safety neurons, then verifies their functional correlation with cross-lingual safety behavior via activation suppression and activation state transfer. Based on this, the authors propose the SK-Masked LoRA defense fine-tuning method, which restricts LoRA adaptation to the FFN output dimensions corresponding to SK, concentrating safety-related changes within the SK range.
Generated by AI from reports · Updated 24 days ago
Attention and developments
Attention, hourly
Now 30, 0% in 24 hours·Peak 30 (Nov 6 03:00)
Related research
This event's paper
- Safety Kernel: Mechanistic Alignment for Cross-Lingual Defense in LLMs
arXiv · 2026-10-17 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesSafety Kernel: Mechanism Alignment for Cross-Lingual Defense of LLMs
The paper proposes Safety Kernel (SK), which locates a set of cross-lingually shared safety neurons through neuron-level cross-lingual safety analysis, and on this basis proposes the SK-Masked LoRA defense fine-tuning method. The authors use multilingual parallel safe and unsafe prompts to compute SCAR scores, identify safety neuron candidates for each language and take their intersection to obtain SK, then verify its functional correlation with cross-lingual safety behavior through activation suppression and activation state transfer. SK-Masked LoRA restricts LoRA adaptation to the FFN output dimensions corresponding to SK, concentrating safety-related changes on cross-lingually shared neuron coordinates. Experiments show that the method greatly reduces attack success rates across multiple languages, including languages not involved in SK discovery, and that it intervenes in only a small number of internal neuron coordinates, with overall limited impact on the model's general capabilities.