Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use

• URL: http://arxiv.org/abs/2603.03205v1
• Date: Tue, 03 Mar 2026 17:59:35 GMT
• Title: Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use
• Authors: Aradhye Agarwal, Gurdit Siyan, Yash Pandya, Joykirat Singh, Akshay Nambi, Ahmed Awadallah,
• Abstract summary: Agentic language models must plan, call tools, and execute long-horizon actions where a single misstep can cause irreversible harm.<n>We introduce MOSAIC, a framework that aligns agents for safe multi-step tool use by making safety decisions explicit and learnable.<n>We show that MOSAIC reduces harmful behavior by up to 50%, increases harmful-task refusal by over 20% on injection attacks, cuts privacy leakage, and preserves or improves benign task performance.
• Score: 6.622648583261088
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Agentic language models operate in a fundamentally different safety regime than chat models: they must plan, call tools, and execute long-horizon actions where a single misstep, such as accessing files or entering credentials, can cause irreversible harm. Existing alignment methods, largely optimized for static generation and task completion, break down in these settings due to sequential decision-making, adversarial tool feedback, and overconfident intermediate reasoning. We introduce MOSAIC, a post-training framework that aligns agents for safe multi-step tool use by making safety decisions explicit and learnable. MOSAIC structures inference as a plan, check, then act or refuse loop, with explicit safety reasoning and refusal as first-class actions. To train without trajectory-level labels, we use preference-based reinforcement learning with pairwise trajectory comparisons, which captures safety distinctions often missed by scalar rewards. We evaluate MOSAIC zero-shot across three model families, Qwen2.5-7B, Qwen3-4B-Thinking, and Phi-4, and across out-of-distribution benchmarks spanning harmful tasks, prompt injection, benign tool use, and cross-domain privacy leakage. MOSAIC reduces harmful behavior by up to 50%, increases harmful-task refusal by over 20% on injection attacks, cuts privacy leakage, and preserves or improves benign task performance, demonstrating robust generalization across models, domains, and agentic settings.

Related papers

• Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models [62.16655896700062]
  Activation steering is a technique to enhance the utility of Large Language Models (LLMs)<n>We show that it unintentionally introduces critical and under-explored safety risks.<n>Experiments reveal that these interventions act as a force multiplier, creating new vulnerabilities to jailbreaks and increasing attack success rates to over 80% on standard benchmarks.
  arXiv  Detail & Related papers  (2026-02-03T12:32:35Z)
• ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack [52.17935054046577]
  We present ReasAlign, a model-level solution to improve safety alignment against indirect prompt injection attacks.<n>ReasAlign incorporates structured reasoning steps to analyze user queries, detect conflicting instructions, and preserve the continuity of the user's intended tasks.
  arXiv  Detail & Related papers  (2026-01-15T08:23:38Z)
• RoboSafe: Safeguarding Embodied Agents via Executable Safety Logic [56.38397499463889]
  Embodied agents powered by vision-language models (VLMs) are increasingly capable of executing complex real-world tasks.<n>However, they remain vulnerable to hazardous instructions that may trigger unsafe behaviors.<n>We propose RoboSafe, a runtime safeguard for embodied agents through executable predicate-based safety logic.
  arXiv  Detail & Related papers  (2025-12-24T15:01:26Z)
• Check Yourself Before You Wreck Yourself: Selectively Quitting Improves LLM Agent Safety [2.7030665672026846]
  Large Language Model (LLM) agents increasingly operate in complex environments with real-world consequences.<n>We propose using "quitting" as a simple yet effective behavioral mechanism for LLM agents to recognize and withdraw from situations where they lack confidence.
  arXiv  Detail & Related papers  (2025-10-18T13:22:19Z)
• Building a Foundational Guardrail for General Agentic Systems via Synthetic Data [76.18834864749606]
  LLM agents can plan multi-step tasks, intervening at the planning stage-before any action is executed-is often the safest way to prevent harm.<n>Existing guardrails mostly operate post-execution, which is difficult to scale and leaves little room for controllable supervision at the plan level.<n>We introduce AuraGen, a controllable engine that synthesizes benign trajectories, injects category-labeled risks with difficulty, and filters outputs via an automated reward model.
  arXiv  Detail & Related papers  (2025-10-10T18:42:32Z)
• LatentGuard: Controllable Latent Steering for Robust Refusal of Attacks and Reliable Response Generation [4.29885665563186]
  LATENTGUARD is a framework that combines behavioral alignment with supervised latent space control for interpretable and precise safety steering.<n>Our results show significant improvements in both safety controllability and response interpretability without compromising utility.
  arXiv  Detail & Related papers  (2025-09-24T07:31:54Z)
• Safety Alignment Can Be Not Superficial With Explicit Safety Signals [8.297367440457508]
  Recent studies on the safety alignment of large language models (LLMs) have revealed that existing approaches often operate superficially.<n>This paper identifies a fundamental cause of this superficiality: existing alignment approaches presume that models can implicitly learn a safety-related reasoning task during the alignment process.<n>By explicitly introducing a safety-related binary classification task and integrating its signals with our attention and decoding strategies, we eliminate this ambiguity.
  arXiv  Detail & Related papers  (2025-05-19T20:40:46Z)
• What Makes and Breaks Safety Fine-tuning? A Mechanistic Study [64.9691741899956]
  Safety fine-tuning helps align Large Language Models (LLMs) with human preferences for their safe deployment. We design a synthetic data generation framework that captures salient aspects of an unsafe input. Using this, we investigate three well-known safety fine-tuning methods.
  arXiv  Detail & Related papers  (2024-07-14T16:12:57Z)
• Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training [67.30423823744506]
  We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position.<n>DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence.
  arXiv  Detail & Related papers  (2024-07-12T09:36:33Z)
• Safe MDP Planning by Learning Temporal Patterns of Undesirable Trajectories and Averting Negative Side Effects [27.41101006357176]
  In safe MDP planning, a cost function based on the current state and action is often used to specify safety aspects. operating based on an incomplete model can often produce unintended negative side effects (NSEs)
  arXiv  Detail & Related papers  (2023-04-06T14:03:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.