Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·27d ago·5 min read
HazardAuditor: From Executable Threats to Safer Computer-Use Agents
TL;DR
HazardAuditor is a new framework designed to enhance the safety of computer-use agents by normalizing their interactions across different platforms. It improves safety decision-making accuracy by up to 16.5 percentage points compared to previous models.
✦ Why It Matters
Implement HazardAuditor to improve the safety of your computer-use agents today.
Key Takeaways
- Utilize HazardAuditor for enhanced safety in computer-use agents.
- Implement Guard Policy Optimization to improve decision-making processes.
- Run agents in controlled environments for better supervision.
- Normalize interactions across frameworks to streamline safety evaluations.
How It Works
HazardAuditor operates by executing various agents in controlled environments, capturing their interactions, and converting these into a canonical event representation. This standardization allows for effective cross-framework supervision, enabling the model to learn from diverse agent behaviors.
Guard Policy Optimization further refines this process by transforming deterministic safety outcomes into sequence-level advantages, ensuring that safety decisions are prioritized during training.