Running AI Agents Inside TEEs Without Losing Your Mind
summary
Architecture for isolating agent-to-agent communication uses confidential VMs, measured boot, attestation nonces and sealed keys, and highlights practical gotchas in key management and session state when agents dynamically invoke unknown tools. Real GPU numbers show 0–7% overhead for large-model LLM inference but up to 45–70% slowdown for short-request workloads; NVLink between GPUs remains a gap—NVLink traffic unprotected—forcing slow CPU routing or weaker guarantees. Apple PCC and Anthropic prototypes illustrate deployment tradeoffs.
AI-written summary · about
chapters (12)
- 0:00AI agents in TEEs
- 2:46What is a TEE
- 5:52Trusted computing base role
- 8:42TEE authentication mechanisms
- 13:53TEE vendor solutions overview
- 17:19AI agents in TEEs features
- 19:22GPU security issues in TEEs
- 21:52Performance overhead in TEEs
- 23:53Apple Private Cloud Compute innovations
- 25:58Anthropic confidential inference
- 27:19TEE threat model limitations
- 28:19Q&A
related talks
20mOwn your agent
24mAI Agents in Production: Why the Platform Matters
01h 42mBuilding Responsible AI Agents in Decentralized Open-Source Environments
34mOpen Source AI Agents on User-Owned Infra: K3s, MCP, & GPU Sharing in Practice
28mThe Hard Truth About AI Agents
06mThe Building Blocks of Agentic Al
18mYour Coding Agent Should Do AI System Engineering
31mBringing and Running AI Agents in the Browser