From Perimeter to Silicon: Why Agentic AI Demands a New Security Model
summary
Agentic AI shifts threats from binary tampering to behavioral and cognitive attacks, so the proposal is to anchor trust with cryptographic attestation and ephemeral agent identities inside confidential VMs. A repeatable pattern decrypts models, prompts, secrets and KV caches just-in-time via a key broker service that verifies measured CPU+GPU attestations; a composite attestation format draft is in IETF. Confidential computing covers roughly 20–30% of controls but provides auditable logs, protected orchestration, and a deployable foundation for governance.
AI-written summary · about
chapters (9)
- 0:00Agentic AI threat models
- 1:30Behavioral threats in AI systems
- 4:41Data in mind and confidential computing
- 6:05Confidential computing across AI stack
- 9:16End-to-end confidential compute for agents
- 12:26CCC’s role in confidential computing
- 17:45Unified attestation for CPU and GPU
- 21:00Future of confidential Agentic AI
- 22:25Q&A
related talks
41mAgentic AI, confidential computing and attested TLS
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
40mIn AI We Trust? Securing the Future, One Agent at a Time
37mConfidential AI: Enabling Trustworthy Ecosystems in the Age of Agentic Intelligence - Session
20mSecurity in the Age of Agentic AI
19mHow Can We Have Private Access to Intelligence?
16mThe Exploding Attack Surface of Agentic AI | F5 and
27mSecurity by design for agentic AI: How CISOs stay in control of autonomous systems