100 Harmless Samples Are All You Need to Break a Safety-Aligned LLM

Think your "clean" fine-tuning data is safe? Recent research shows that just 100 samples from benign datasets like Dolly and Alpaca — innocent on the surface, yet carrying harmful gradient directions — are enough to collapse the safety alignment of an already-aligned LLM. And it works anchor-free: no external malicious data required. In this talk, we put Twinkle AI's open-source gemma-3-4B-T1-it on the table for a hands-on attack-and-defense walkthrough: how to surface these "invisible landmines" from Traditional Chinese benign datasets using Self-Inf-N, why they're so hard to undo in continual learning scenarios, and whether gradient-constrained defenses like SafeGrad and AsFT can actually hold the line. For developers and researchers interested in LLM safety, fine-tuning, and open-source model governance.