# Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue

arXiv: https://arxiv.org/abs/2609.36849 · HTML: https://mlsanity.com/paper/2609.36849/

## 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 28 references that are in the map.
- First posted on arXiv 29 Sep 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 / [Jailbreaking](https://mlsanity.com/safety/jailbreaking/)

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

- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection](https://mlsanity.com/paper/2302.12173/) (2023)
- [Jailbreaking Black Box Large Language Models in Twenty Queries](https://mlsanity.com/paper/2310.08419/) (2023)
- [Qwen2.5-Coder Technical Report](https://mlsanity.com/paper/2409.12186/) (2024)
- [Tree of Attacks: Jailbreaking Black-Box LLMs Automatically](https://mlsanity.com/paper/2312.02119/) (2023)
- [Baseline Defenses for Adversarial Attacks Against Aligned Language Models](https://mlsanity.com/paper/2309.00614/) (2023)
- [WildChat: 1M ChatGPT Interaction Logs in the Wild](https://mlsanity.com/paper/2405.01470/) (2024)
- [SmoothLLM: Defending Large Language Models Against Jailbreaking Attacks](https://mlsanity.com/paper/2310.03684/) (2023)
- [Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack](https://mlsanity.com/paper/2404.01833/) (2024)
- [LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet](https://mlsanity.com/paper/2408.15221/) (2024)
- [GradSafe: Detecting Jailbreak Prompts for LLMs via Safety-Critical Gradient Analysis](https://mlsanity.com/paper/2402.13494/) (2024)
- [Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes](https://mlsanity.com/paper/2403.00867/) (2024)
- [Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning](https://mlsanity.com/paper/2408.09600/) (2024)
- [Vulnerability-Aware Alignment: Mitigating Uneven Forgetting in Harmful Fine-Tuning](https://mlsanity.com/paper/2506.03850/) (2025)
- [Beware What You Autocomplete: Forensic Attribution of Backdoored Code Completions](https://mlsanity.com/paper/2607.08011/) (2026)
- [NeuronGuard: Robust LLM Safety Alignment via Ablation-Aware Safety Signal Redistribution](https://mlsanity.com/paper/2608.23959/) (2026)

## Abstract (arXiv)

Safety-aligned language models are commonly deployed as multi-turn assistants, which lets adversaries spread unsafe intent across several user turns instead of a single prompt. Gradient-based jailbreak detectors such as GradSafe were developed for single prompts: they score an input by the alignment between its induced gradient and a fixed unsafe reference direction, and their effectiveness in multi-turn dialogue remains unclear. We conduct a controlled evaluation of gradient-based jailbreak detection in multi-turn settings. We extend GradSafe with a Context Window Scanner that applies the detector to fixed-size windows of user turns and uses the maximum window score as the conversation-level score. We evaluate different window sizes, attack families, benign conversation distributions, and target models. The results differ sharply between synthetic and realistic benign settings. Against synthetic benign conversations, the detector achieves an ROC-AUC of 0.98 on human-authored multi-turn jailbreaks. On WildChat benign conversations, ROC-AUC drops to 0.76, and a threshold calibrated on synthetic data flags more than 90% of benign conversations as unsafe. Under realistic benign distributions, single-turn windows give the highest separability, whereas longer windows and accumulated contexts reduce performance. The detector is also sensitive to the attack-generation method and target model: successful Crescendo attacks receive scores comparable to or lower than benign conversations, and Qwen2.5-7B-Instruct yields near-random separability with a different optimal window size. These findings show that gradient-based signals can support multi-turn jailbreak detection, but reliable deployment requires calibration on realistic benign conversations, short-window scoring, length-aware thresholds, and evaluation across attack types and model architectures.