Safety evaluation of large language models (LLMs) increasingly relies on LLM-as-a-Judge frameworks, but the high cost of frontier models limits scalability. We propose a cost-efficient multi-agent judging framework that employs Small Language Models (SLMs) through structured debates among critic, defender, and judge agents. To rigorously assess safety judgments, we construct HAJailBench, a large-scale human-annotated jailbreak benchmark comprising 12,000 adversarial interactions across diverse attack methods and target models. The dataset provides fine-grained, expert-labeled ground truth for evaluating both safety robustness and judge reliability. Our SLM-based framework achieves agreement comparable to GPT-4o judges on HAJailBench while substantially reducing inference cost. Ablation results show that three rounds of debate yield the optimal balance between accuracy and efficiency. These findings demonstrate that structured, value-aligned debate enables SLMs to capture semantic nuances of jailbreak attacks and that HAJailBench offers a reliable foundation for scalable LLM safety evaluation.

Ensuring Large Language Model (LLM) safety remains challenging due to the absence of universal standards and reliable content validators, making it difficult to obtain effective training signals. We discover that aligned models already possess robust internal safety beliefs: they consistently produce high-confidence refusals to harmful requests while exhibiting high entropy when generating potentially dangerous content. This entropy gap reveals an untapped signal--models intrinsically "know" when to refuse. We introduce Safety Instincts Reinforcement Learning (SIRL), which transforms this internal confidence into a self-generated reward signal, eliminating dependence on external validators or human annotations. SIRL teaches models to trust their safety instincts by reinforcing low-entropy refusal behaviors. Evaluated on Llama and Qwen models, SIRL maintains 89%+ Defense Success Rates (DSRs) against 20+ jailbreak methods, from static prompts to adaptive attacks. Using only 15,000 unlabeled prompts, SIRL surpasses resource-intensive supervised methods while preserving performance on mathematics, coding, and conversation benchmarks. Our work demonstrates that effective alignment can emerge from within, paving the way for more autonomous and robust AI safety mechanisms that scale without extensive human oversight.

The remarkable capabilities of Large Language Models (LLMs) have raised significant safety concerns, particularly regarding "jailbreak" attacks that exploit adversarial prompts to bypass safety alignment mechanisms. Existing defense research primarily focuses on single-turn attacks, whereas multi-turn jailbreak attacks progressively break through safeguards through by concealing malicious intent and tactical manipulation, ultimately rendering conventional single-turn defenses ineffective. To address this critical challenge, we propose the Bidirectional Intention Inference Defense (BIID). The method integrates forward request-based intention inference with backward response-based intention retrospection, establishing a bidirectional synergy mechanism to detect risks concealed within seemingly benign inputs, thereby constructing a more robust guardrails that effectively prevents harmful content generation. The proposed method undergoes systematic evaluation compared with a no-defense baseline and seven representative defense methods across three LLMs and two safety benchmarks under 10 different attack methods. Experimental results demonstrate that the proposed method significantly reduces the Attack Success Rate (ASR) across both single-turn and multi-turn jailbreak attempts, outperforming all existing baseline methods while effectively maintaining practical utility. Notably, comparative experiments across three multi-turn safety datasets further validate the proposed model's significant advantages over other defense approaches.

While fine-tuning services drive the rapid expansion of task capabilities in large language models (LLMs), they are often accompanied by the degradation and reorganization of safety-aligned representations, making models more prone to deviating from human preferences and exposing them to emerging jailbreak risks. Existing post-fine-tuning defense methods predominantly rely on single-scale safety correction mechanisms, which struggle to achieve a robust balance among safety, model utility, and continual adaptability. We propose Multi-Level Safety Continual Projection (MSCP), a training-free post-fine-tuning safety enhancement method that implicitly aligns global and localized safety activations through coordinated multi-level representations to isolate sparse neuron clusters governing safety-sensitive behaviors. It then applies composable safety-direction projections without retraining, effectively suppressing harmful outputs under minimal parameter perturbations while preserving task performance and improving alignment with human preferences. Extensive experiments across multiple fine-tuned LLM models demonstrate that our method significantly reduce harmfulness scores and attack success rates with minimal parameter modifications, while preserving the model's utility. Furthermore, we introduce a task-specific, multi-dimensional heterogeneous safety activation clustering mechanism that enables continual defense and generalization capability against unforeseen emerging safety concerns.