# The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play

arXiv: https://arxiv.org/abs/2605.08427 · HTML: https://mlsanity.com/paper/2605.08427/

## Key facts

- Cited by 1 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 41 references that are in the map.
- First posted on arXiv 8 May 2026.

## Placed in

- LLM reasoning / [Reasoning for agents, planning and code](https://mlsanity.com/lr/reasoning-for-agents-planning-and-code/)
- LLM safety and security / [Prompt injection and agent security](https://mlsanity.com/safety/prompt-injection-and-agent-security/)

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

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [QLoRA: Efficient Finetuning of Quantized LLMs](https://mlsanity.com/paper/2305.14314/) (2023)
- [Trust Region Policy Optimization](https://mlsanity.com/paper/1502.05477/) (2015)
- [BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions](https://mlsanity.com/paper/1905.10044/) (2019)
- [CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge](https://mlsanity.com/paper/1811.00937/) (2018)
- [Red Teaming Language Models with Language Models](https://mlsanity.com/paper/2202.03286/) (2022)
- [Tree of Attacks: Jailbreaking Black-Box LLMs Automatically](https://mlsanity.com/paper/2312.02119/) (2023)
- [Safe RLHF: Safe Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2310.12773/) (2023)
- [Baseline Defenses for Adversarial Attacks Against Aligned Language Models](https://mlsanity.com/paper/2309.00614/) (2023)
- [WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs](https://mlsanity.com/paper/2406.18495/) (2024)
- [REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization](https://mlsanity.com/paper/2501.03262/) (2025)
- [WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models](https://mlsanity.com/paper/2406.18510/) (2024)
- [Efficient Adversarial Training in LLMs with Continuous Attacks](https://mlsanity.com/paper/2405.15589/) (2024)
- [A Comprehensive Study of Jailbreak Attack versus Defense for Large Language Models](https://mlsanity.com/paper/2402.13457/) (2024)
- [Curiosity-driven Red-teaming for Large Language Models](https://mlsanity.com/paper/2402.19464/) (2024)
- [Open Challenges in Multi-Agent Security: Towards Secure Systems of Interacting AI Agents](https://mlsanity.com/paper/2505.02077/) (2025)
- [Adversarial Attacks and Defenses in Large Language Models: Old and New Threats](https://mlsanity.com/paper/2310.19737/) (2023)
- [SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation](https://mlsanity.com/paper/2501.01765/) (2025)
- [Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models](https://mlsanity.com/paper/2506.07468/) (2025)
- [Explaining Predictions of Non-Linear Classifiers in NLP](https://mlsanity.com/paper/1606.07298/) (2016)
- [MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety](https://mlsanity.com/paper/2602.01539/) (2026)
- [Global Convergence of Localized Policy Iteration in Networked Multi-Agent Reinforcement Learning](https://mlsanity.com/paper/2211.17116/) (2022)

## Built on by (most cited first, 1 of 1)

- [Reflections and Fragments: Securing LLMs Against Sequential Mosaic Attacks](https://mlsanity.com/paper/2610.05346/) (2026)

## Abstract (arXiv)

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.