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

TL;DR

This paper introduces MOSAIC, a framework that explicitly aligns agentic language models for safe multi-step tool use by making safety decisions explicit, improving safety and robustness across various tasks and models.

Contribution

MOSAIC is a novel post-training alignment method that structures safety as explicit, learnable decisions within an inference loop, addressing limitations of existing methods in agentic settings.

Findings

Reduces harmful behavior by up to 50%

Increases harmful-task refusal by over 20% on injection attacks

Cuts privacy leakage while maintaining or improving benign task performance

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.