Enterprise AI Analysis
Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use
Explore MOSAIC: a novel framework empowering AI agents with explicit safety reasoning and refusal, ensuring robust and reliable multi-step tool use.
Executive Impact
MOSAIC introduces a post-training framework that enables AI agents to make explicit, learnable safety decisions during multi-step tool use. By structuring inference as a plan, check, then act or refuse loop and utilizing preference-based reinforcement learning, MOSAIC significantly reduces harmful behavior, increases refusal rates for dangerous tasks, and cuts privacy leakage across various models and domains. This approach demonstrates that agentic safety is driven by structured inference and temporal safety decisions rather than merely model scale.
50%
Harm Reduction (Qwen2.5)
20%
Refusal Increase (Injection Attacks)
23%
Privacy Leakage Cut
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
MOSAIC Overview
Training Methodology
Performance Results
Modular Reasoning for Agentic Safety
MOSAIC organizes agentic reasoning as a plan, check, then act or refuse loop. This modular design makes safety decisions explicit and learnable, allowing agents to dynamically assess risks, handle adversarial tool feedback, and manage overconfident intermediate reasoning. It's a post-training framework that aligns agents for safe multi-step tool use.
MOSAIC Framework Process
User gives the task
→
Plan (propose tool)
→
Implicit check / Optional safety check (reasoning)
→
Act (tool execution) / Refuse
→
Report task completion or refusal
Preference-Based Reinforcement Learning
MOSAIC employs preference-based reinforcement fine-tuning with pairwise trajectory comparisons. This method captures safety distinctions often missed by scalar rewards, such as preferring early refusal over late aborts. It optimizes policies by jointly balancing safety alignment, task utility, structured outputs, and token efficiency, without relying on trajectory-level labels.
| Feature | Traditional LLM Agents | MOSAIC Agents |
|---|---|---|
| Safety Decision |
|
|
| Refusal Mechanism |
|
|
| Training Data |
|
|
| Handling Harmful Tasks |
|
|
| Context Overhead |
|
|
| Generalization |
|
|
Model-Adaptive Safety & Utility Gains
MOSAIC demonstrates significant model-adaptive gains across Qwen2.5-7B, Qwen3-4B-Thinking, and Phi-4. It reduces harmful behavior by up to 50%, boosts benign task completion by 93% by avoiding reasoning loops, and reduces over-refusal by 56% for conservative models. These gains come with minimal overhead and robust generalization.
Case Study: Qwen2.5-7B Safety Hardening
On AgentHarm, MOSAIC reduced Qwen2.5's harmful-task score by 50% (from 0.18 to 0.09) and increased harmful-task refusal from 0.74 to 0.87. This shows substantial safety gains with limited utility loss, improving robustness against prompt injection attacks.
Key Highlight: 50% reduction in harmful-task score.
Calculate Your Potential ROI
Estimate the tangible benefits of integrating advanced AI safety frameworks into your enterprise operations.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Your Implementation Roadmap
A phased approach to integrating MOSAIC into your enterprise, ensuring a smooth and secure transition.
Phase 1: Assessment & Strategy
Conduct a comprehensive audit of existing AI systems, identify high-risk agentic workflows, and define custom safety policies tailored to your operational needs.
Phase 2: MOSAIC Integration & Training
Implement the MOSAIC framework, fine-tune models using preference-based RL, and integrate explicit safety checks into your agent's decision loops. Pilot with non-critical workflows.
Phase 3: Rollout & Continuous Optimization
Gradual deployment across enterprise, continuous monitoring of safety metrics, and iterative refinement of agent behavior based on real-world feedback and emerging threats.
Ready to Implement MOSAIC?
Schedule a personalized consultation with our AI experts to discover how MOSAIC can elevate your enterprise's agentic safety and performance.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat