Panel Discussion: State of the Union: Confidential AI From the Field
summary
Panelists report concrete wins and remaining gaps: a UAE mandate pushes 50% mandate for autonomous agents within two years; SWIFT runs model validation across 450+ financial orgs inside confidential enclaves; healthcare uses combined personal and population records for personalized medicine; practical deployments show negligible performance tax and agent control layers running in TE while large LLMs remain self‑hosted in confidential GPUs; hard problems are composing attestations across CPUs/GPUs/DPUs and preventing KV cache leaks where 95% of a session can reside in cache.
AI-written summary · about
chapters (18)
- 0:00Panel intro and confidential AI basics
- 1:04Defining confidential AI
- 3:11UAE AI agent mandates
- 5:18Model weight protection focus
- 7:28Confidentiality improves AI conversations
- 9:11Silicon vendor challenges
- 12:00Verifiable AI workloads
- 14:09Personalized medicine driver
- 15:55Military intelligence sharing
- 18:43Verifying agent behavior
- 20:05Assumptions disproven in 12 months
- 21:32Integrity in agentic AI
- 23:00Verifiability across cloud/on-prem
- 25:06Attestation across AI pipeline
- 27:15Portable governance records
- 27:38Attestation-trust gap in production
- 29:05Multi-tenancy privacy risks
- 30:05Summary
related talks
32mPanel Discussion: Trends in Generative AI
24mAI, Health & Confidentiality
37mConfidential AI: Enabling Trustworthy Ecosystems in the Age of Agentic Intelligence - Session
36mCTO Panel: The Impact of Confidential Computing with Intel, Azure, AMD & NVIDIA
31mPanel: Trustworthy AI: Enhancing Human Agency and Machine Accountability in the Age of Agentic AI
22mFireside Chat - AI Security Vulnerabilities
13mProvably Secure AI for Health Data – Powered by Confidential Computing
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…