AI 速览：作者提出 Assess→Decide→Respond 框架，让防御方在多轮对话的关键轮次先显式做出可解析的安全决策，再据此生成回复，以利用对齐的“浅层性”进行防御。
Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess $\to$ Decide $\to$ Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.
Chengquan Guo, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Jialin Song, Qiyao Ma, Carolyn Buractaon, Christopher White 等 10 位作者
LLM agents are increasingly adopted in long-horizon tasks and interact with diverse external environments. Despite their growing capabilities, these agents remain vulnerable to real-world attacks such as indirect prompt injections, which manipulate agents through malicious instructions embedded in external environments. While existing risk assessment work has primarily focused on the injection content, the effect of where an injection appears has not been systematically characterized. Studying injection placement requires agents to operate over longer-horizon tasks and interact with diverse environmental surfaces. Thus, we build three agentic testbeds spanning tool use, coding, and daily workflow domains with diverse injection surfaces. Holding the injection content fixed, we find that agent susceptibility varies across injection placements. We refer to this phenomenon as the trust region effect: agents implicitly assign different levels of trust to different regions of their environments. For example, an agent powered by GPT-5.4 follows an unsafe injection instruction placed in a repository documentation file (i.e., a README or TODO file) 2.6 times more often on average than the same instruction placed in a code or data file. Building upon this observation, we develop learning-based attackers that generate injection content and select better injection placements. Our approach demonstrates that both where an injection is placed and what content it contains can affect the strength of indirect prompt injection attacks. To the best of our knowledge, this work provides the first systematic characterization for the effectiveness of different injection placements, spanning both coarse-grained environmental surfaces and fine-grained positions. The uneven trust that agents implicitly place across their environments provides important insights for future attack and defense design.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but losing sight of the victim's replies, wasting turns and limiting transferability; interactive generators read each victim reply to generate the next-turn attack but require in-loop evaluators for success signals or hints, incurring substantial costs and inflated performance. Bridging this gap, we propose SEMA-Auto, a sighted and efficient multi-turn attacker with autonomy, interacting with the victim model and terminating the session by itself upon success. To stabilize its behaviors and boost its capabilities, SEMA-Auto is trained in two stages. Synthetic bootstrap tuning installs the notion of attack targets and conforming decisions by fine-tuning the base model with bootstrapped trajectories. Decision-supervised reinforcement learning then optimizes two objectives jointly yet separately, reinforcing the attack with higher jailbreak reward while supervising the decision at each turn against an absolute target. Our method achieves state-of-the-art (SOTA) attack success rates (ASR) across multiple datasets, victims, and judges, outperforming all static and interactive baselines, within almost the fewest turns, demonstrating both effectiveness and efficiency. For example, SEMA-Auto performs a $76.0\%$ ASR@1 within an average of $3.0$ turns against DeepSeek-V4-Flash on AdvBench, while the best static baseline is $54.8\%$ with fixed $7$ turns and the best interactive baseline is $43.8\%$ in $3.7$ turns. Against safety-hardened frontier models, GPT-oss-120B, GPT-5.4, Claude-Haiku-4.5, and GPT-6-Astra, our counter-trained attacker achieves an average $59.6\%$ ASR@1 on HarmBench, $13.8\%$ over prior SOTA. Our code is available at: https://github.com/microsoft/SEMA