Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 11·5 min read
A Five-Plane Reference Architecture for Runtime Governance of Production AI Agents
TL;DR
AI agents in production often lack effective governance, leading to risks in decision-making and compliance. A five-plane reference architecture was developed to provide runtime governance for these agents, focusing on aspects like monitoring and accountability.
✦ Why It Matters
Engineers can implement this architecture to improve AI governance, ensuring compliance and reducing operational risks.
Key Takeaways
- Adopt the five-plane architecture for AI governance.
- Utilize stateful evaluation for complex workflows.
- Implement stop-anywhere mediation for flexible control.
- Ensure compliance with structured audit evidence.
How It Works
The architecture operates by decomposing governance into five distinct planes. The reasoning plane assesses the intent behind actions, while the four enforcement planes apply security measures based on identity, network, endpoint, and data considerations.
This layered approach allows for real-time decision-making and the ability to manage risks associated with delegated actions effectively.