Organizations: Stony Brook University · Amazon Nova Responsible AI
Abstract
Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical vulnerability: an imperfect Reward Model (RM) can become a single point of failure when it fails to penalize unsafe behaviors. While existing red-teaming approaches primarily target policy-level weaknesses, they overlook what we term systemic weaknesses cases where both the core LLM and the RM fail in tandem. We present ARES, a framework that systematically discovers and mitigates such dual vulnerabilities. ARES employs a ``Safety Mentor'' that dynamically composes semantically coherent adversarial prompts by combining structured component types (topics, personas, tactics, goals) and generates corresponding malicious and safe responses. This dual-targeting approach exposes weaknesses in both the core LLM and the RM simultaneously. Using the vulnerabilities gained, ARES implements a two-stage repair process: first fine-tuning the RM to better detect harmful content, then leveraging the improved RM to optimize the core model. Experiments across multiple adversarial safety benchmarks demonstrate that ARES substantially enhances safety robustness while preserving model capabilities, establishing a new paradigm for comprehensive RLHF safety alignment.
Despite advances in safety alignment, prompt-rewriting attacks such as persona modulation, fictional framing and persuasion-based reformulation, can bypass safety filters even on frontier models. Existing defenses either rely on non-scalable human curation or white-box optimisation that overfits to specific model internals, leaving aligned models brittle against the very class of adaptive black-box adversaries they will face in deployment. To address this gap, we introduce CHASE (Co-evolutionary Hardening through Adversarial Safety-Escalation), a closed-loop red-blue teaming framework in which a black-box attacker and a safety-aligned defender co-evolve. The attacker is trained via Group Relative Policy Optimization (GRPO) under a multiplicative reward that jointly enforces bypass effectiveness and intent fidelity, while the defender is hardened on the harvested adversarial rewrites through a two-stage GRPO + rejection-sampled SFT pipeline balanced with benign data. Evaluated on BeaverTails and JailbreakBench against five held-out attack families (PAIR, TAP, AutoDAN, PAP, Translation), CHASE cuts mean StrongREJECT score by 43.2% with 0% false-refusal on benign prompts. Beyond the headline result, CHASE shows that template-free RL exploration recovers latent attack primitives that transfer across mechanistically distinct attack families, suggesting a path toward LLM safety hardening that generalises beyond the narrow distributions achieved thus far in adversarial training.
Rahul Markasserithodi, Aditya Joshi, Yuekang Li +3
Safety-aligned large language models rely on RLHF and instruction tuning to refuse harmful requests, yet the internal mechanisms implementing safety behavior remain poorly understood. We introduce the Attention Redistribution Attack (ARA), a white-box adversarial attack that identifies safety-critical attention heads and crafts nonsemantic adversarial tokens that redirect attention away from safety-relevant positions. Unlike prior jailbreak methods operating at the semantic or output-logit level, ARA targets the geometry of softmax attention on the probability simplex using Gumbel-softmax optimization over targeted heads. Across LLaMA-3-8B-Instruct, Mistral-7B-Instruct-v0.1, and Gemma-2-9B-it, ARA bypasses safety alignment with as few as 5 tokens and 500 optimization steps, achieving 36% ASR on Mistral-7B and 30% on LLaMA-3 against 200 HarmBench prompts, while Gemma-2 remains at 1%. Our principal mechanistic finding is a dissociation between ablation and redistribution: zeroing out the top-ranked safety heads produces at most 1 flip among 39 to 50 baseline refusals, while ARA targeting the corresponding safety-heavy layers flips 72/200 prompts on Mistral-7B and 60/200 on LLaMA-3. This suggests that safety is not localized in these heads as removable components, but emerges from the attention routing they perform. Removing a head allows compensation through the residual stream, while redirecting its attention propagates a corrupted signal downstream.
Conventional large language model (LLM) safety alignment relies on a reactive, disjoint loop: attackers exploit a static model, then defenders patch exposed vulnerabilities. This sequential setup leads to attackers overfitting obsolete exploits while defenders perpetually lag behind emerging threats. To address this, we introduce Self-RedTeam, the first fully online self-play multi-agent reinforcement learning (MARL) algorithm that continuously co-evolves attacker and defender for robust safety alignment. A single policy self-plays as both attacker and defender, generating adversarial prompts and defending against them, with a reward model adjudicating outcomes. Each role uses hidden chain-of-thought for strategic planning. Grounded in two-player zero-sum game theory, we establish a theoretical safety guarantee: if the game converges to Nash Equilibrium, the defender produces safe responses against any adversarial input. Empirically, Self-RedTeam generalizes across five models from the Llama and Qwen families, uncovering more diverse attacks (+17.80% SBERT) and improving safety of RLHF-trained models by up to 95% across 14 benchmarks. Our work motivates a shift from reactive patching to proactive co-evolution, enabling LLM safety self-improvement via online self-play MARL. Link to code: https://github.com/mickelliu/selfplay-redteaming