# When Autoregressive Consistency Hurts Safety Alignment

arXiv: https://arxiv.org/abs/2606.04168 · HTML: https://mlsanity.com/paper/2606.04168/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 16 of its 24 references that are in the map.
- First posted on arXiv 2 Jun 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 16 of 16)

- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Towards Deep Learning Models Resistant to Adversarial Attacks](https://mlsanity.com/paper/1706.06083/) (2017)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [Jailbreaking Black Box Large Language Models in Twenty Queries](https://mlsanity.com/paper/2310.08419/) (2023)
- [Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://mlsanity.com/paper/2310.03693/) (2023)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models](https://mlsanity.com/paper/2310.04451/) (2023)
- [Tree of Attacks: Jailbreaking Black-Box LLMs Automatically](https://mlsanity.com/paper/2312.02119/) (2023)
- [Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks](https://mlsanity.com/paper/2404.02151/) (2024)
- [Safety Alignment Should Be Made More Than Just a Few Tokens Deep](https://mlsanity.com/paper/2406.05946/) (2024)
- [The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning](https://mlsanity.com/paper/2312.01552/) (2023)
- [SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decoding](https://mlsanity.com/paper/2402.08983/) (2024)
- [Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs](https://mlsanity.com/paper/2407.15549/) (2024)
- [The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence](https://mlsanity.com/paper/2502.17420/) (2025)
- [Defending Against Unforeseen Failure Modes with Latent Adversarial Training](https://mlsanity.com/paper/2403.05030/) (2024)
- [LLMs Encode Harmfulness and Refusal Separately](https://mlsanity.com/paper/2507.11878/) (2025)

## Abstract (arXiv)

Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-tuning mainly reshapes the model's behavior near the first few output tokens. We argue that this phenomenon can be understood through autoregressive consistency, the tendency of next-token prediction to preserve and extend the current response trajectory consistently. By analyzing the learning dynamics of safety alignment, we show that autoregressive consistency can concentrate alignment updates on early tokens, offering a mechanistic explanation for shallow safety alignment. The same mechanism also predicts a broader class of attacks on LLMs: attacks that induce harmful continuation states at arbitrary positions in the output trajectory. As a concrete example, we introduce random insertion attack, which inserts a short harmful span into an otherwise safe refusal trajectory and exploits autoregressive consistency to sustain the resulting harmful branch, thereby bypassing safety alignment. Notably, a short harmful span can redirect the generation to be harmful even after a long refusal prefix, highlighting autoregressive consistency as a potential broader failure mechanism. This suggests that safety alignment should also break harmful autoregressive consistency throughout the output trajectory. We therefore propose adversarial safety alignment, an initial framework based on worst-case harmful continuation states, and instantiate it with random worst-insertion training. Overall, our results suggest that autoregressive consistency should be treated as a central consideration in both safety alignment and attack design.