Fail-Closed Alignment for Large Language Models

TL;DR

This paper introduces fail-closed alignment, a new approach for making large language models safer by ensuring refusal mechanisms remain effective even when some pathways fail, demonstrated through a progressive framework that enhances robustness against jailbreak attacks.

Contribution

The paper proposes fail-closed alignment as a novel safety principle and presents a progressive alignment method that identifies and reinforces independent refusal directions in LLMs.

Findings

Achieves strongest robustness against jailbreaks among tested methods.

Mitigates over-refusal while maintaining generation quality.

Models encode multiple independent refusal directions.

Abstract

We identify a structural weakness in current large language model (LLM) alignment: modern refusal mechanisms are fail-open. While existing approaches encode refusal behaviors across multiple latent features, suppressing a single dominant feature$-$via prompt-based jailbreaks$-$can cause alignment to collapse, leading to unsafe generation. Motivated by this, we propose fail-closed alignment as a design principle for robust LLM safety: refusal mechanisms should remain effective even under partial failures via redundant, independent causal pathways. We present a concrete instantiation of this principle: a progressive alignment framework that iteratively identifies and ablates previously learned refusal directions, forcing the model to reconstruct safety along new, independent subspaces. Across four jailbreak attacks, we achieve the strongest overall robustness while mitigating over-refusal and preserving generation quality, with small computational overhead. Our mechanistic analyses confirm that models trained with our method encode multiple, causally independent refusal directions that prompt-based jailbreaks cannot suppress simultaneously, providing empirical support for fail-closed alignment as a principled foundation for robust LLM safety.