January 22, 2026 - A new study argues that common LLM safety mechanisms remain vulnerable to early-token manipulations and other prompt-level exploits. The paper, “Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism,” suggests that refusal layers are too shallow and can be bypassed with carefully crafted input sequences.

The authors propose probabilistic ablation techniques that remove or dampen unsafe behaviors deeper in the model’s representations, rather than relying solely on surface-level refusal policies. They also recommend evaluation protocols that test adversarial inputs earlier in the token stream.

The findings arrive as regulators and enterprise buyers demand clearer safety guarantees. If the proposed methods are validated at scale, they could influence future alignment strategies for both open and closed models.

Credit: Yuanbo Xie, Yingjie Zhang, Tianyun Liu, and co-authors. Primary source: arXiv:2509.15202.