# When Models Outthink Their Safety: Unveiling and Mitigating Self-Jailbreak in Large Reasoning Models

arXiv: https://arxiv.org/abs/2510.21285 · HTML: https://mlsanity.com/paper/2510.21285/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- Placed in Reasoning for agents, planning and code, one of 17 directions in LLM reasoning.
- Builds on 36 of its 48 references that are in the map.
- First posted on arXiv 24 Oct 2025.

## Placed in

- LLM reasoning / [Reasoning for agents, planning and code](https://mlsanity.com/lr/reasoning-for-agents-planning-and-code/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 30 of 36)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [A Survey on LLM-as-a-Judge](https://mlsanity.com/paper/2411.15594/) (2024)
- [AgentBench: Evaluating LLMs as Agents](https://mlsanity.com/paper/2308.03688/) (2023)
- [OpenAI o1 System Card](https://mlsanity.com/paper/2412.16720/) (2024)
- [DeepSeek-V3 Technical Report](https://mlsanity.com/paper/2412.19437/) (2024)
- [SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild](https://mlsanity.com/paper/2503.18892/) (2025)
- [Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs](https://mlsanity.com/paper/2503.01307/) (2025)
- [Deliberative Alignment: Reasoning Enables Safer Language Models](https://mlsanity.com/paper/2412.16339/) (2024)
- [SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal](https://mlsanity.com/paper/2406.14598/) (2024)
- [WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models](https://mlsanity.com/paper/2406.18510/) (2024)
- [Large Language Model Agent: A Survey on Methodology, Applications and Challenges](https://mlsanity.com/paper/2503.21460/) (2025)
- [Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs](https://mlsanity.com/paper/2308.13387/) (2023)
- [From LLM Reasoning to Autonomous AI Agents: A Comprehensive Review](https://mlsanity.com/paper/2504.19678/) (2025)
- [SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities](https://mlsanity.com/paper/2502.12025/) (2025)
- [Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable](https://mlsanity.com/paper/2503.00555/) (2025)
- [ALERT: A Comprehensive Benchmark for Assessing Large Language Models' Safety through Red Teaming](https://mlsanity.com/paper/2404.08676/) (2024)
- [The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1](https://mlsanity.com/paper/2502.12659/) (2025)
- [The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm](https://mlsanity.com/paper/2406.18682/) (2024)
- [STAR-1: Safer Alignment of Reasoning LLMs with 1K Data](https://mlsanity.com/paper/2504.01903/) (2025)
- [RealSafe-R1: Safety-Aligned DeepSeek-R1 without Compromising Reasoning Capability](https://mlsanity.com/paper/2504.10081/) (2025)
- [Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreaking](https://mlsanity.com/paper/2502.12970/) (2025)
- [Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?](https://mlsanity.com/paper/2502.12206/) (2025)
- [Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers](https://mlsanity.com/paper/2506.15674/) (2025)
- [Unveiling Safety Vulnerabilities of Large Language Models](https://mlsanity.com/paper/2311.04124/) (2023)
- [SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment](https://mlsanity.com/paper/2505.14667/) (2025)
- [To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models](https://mlsanity.com/paper/2502.12202/) (2025)
- [Are Smarter LLMs Safer? Exploring Safety-Reasoning Trade-offs in Prompting and Fine-Tuning](https://mlsanity.com/paper/2502.09673/) (2025)

## Built on by (most cited first, 6 of 6)

- [Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability](https://mlsanity.com/paper/2604.06628/) (2026)
- [Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures](https://mlsanity.com/paper/2601.19928/) (2026)
- [Do Reasoning LLMs Refuse What They Infer in Long Contexts?](https://mlsanity.com/paper/2602.08874/) (2026)
- [Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training](https://mlsanity.com/paper/2605.28467/) (2026)
- [First Token Matters: Understanding Safety Collapse in Large Reasoning Models](https://mlsanity.com/paper/2609.18471/) (2026)
- [PyINE: A Framework for Scalable Elicitation and Oversight via Code Execution](https://mlsanity.com/paper/2610.04737/) (2026)

## Abstract (arXiv)

Large Reasoning Models (LRMs) achieve strong performance on complex multi-step reasoning, yet they still exhibit severe safety failures such as harmful content generation. Existing methods often apply coarse-grained constraints over the entire reasoning trajectories, which can undermine reasoning capability while failing to address the root causes of unsafe behavior. In this work, we uncover a previously underexplored failure mode in LRMs, termed Self-Jailbreak, where models initially recognize the harmful intent of a query, but override this judgment during subsequent reasoning steps, ultimately generating unsafe outputs. Such a phenomenon reveals that LRMs are capable of recognizing harm, while safety failures primarily arise from reasoning steps. Motivated by this finding, we propose Chain-of-Guardrail(CoG), a trajectory-level training framework that mitigates Self-Jailbreak via targeted, step-level interventions while maintaining reasoning ability. Experiments across multiple safety and reasoning benchmarks indicate that CoG achieves a favorable balance between safety and reasoning performance compared with existing approaches.