Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·25d ago·5 min read
Compositional Policy Violations: When Step-Level Compliance Fails In Agentic AI Workflows
TL;DR
Compositional Policy Violations (CPVs) occur when individual steps in AI workflows comply with regulations, but the overall execution does not. A new architecture evaluates policies over complete execution traces to address this issue.
✦ Why It Matters
Implement a provenance-aware architecture to ensure compliance across entire AI workflows today.
Key Takeaways
- Assess compliance at the workflow level, not just individual steps.
- Identify and categorize types of Compositional Policy Violations.
- Implement a system to evaluate policies over complete execution traces.
- Recompute critical quantities from raw data for accurate compliance.
How It Works
The proposed architecture tracks the entire execution of AI workflows, capturing raw provenance data. By analyzing this comprehensive data, it can accurately evaluate compliance with organizational policies, addressing the limitations of step-scoped checks.