Agentic AI, confidential computing and attested TLS
summary
Argues that combining agentic AI with confidential computing requires attestation of model weights as executable code, source code, system prompts, tool lists and invocation order, and GPU state; proposes an attested TLS-enabled MCP server to provide endpoint attestation during inference (not training); notes current TLS solutions are pre- or intra-handshake and do not protect runtime-modified agents, so post-handshake attestation and supply-chain checks (compilers, drivers, accelerators) plus defenses against prompt injection and rogue-agent payment flows are needed.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 10:37Defining agentic AI
- 21:18Training vs inference in AI agents
- 30:08Key threats and security concerns
- 37:11Differences from traditional workloads
- 40:23Threats and future directions
related talks
27mFrom Perimeter to Silicon: Why Agentic AI Demands a New Security Model
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
12mKeynote: Protecting Data in the Agentic AI Era
26mTrust Is the Next Bottleneck: Why the Agentic Economy Needs Confidential
37mConfidential AI: Enabling Trustworthy Ecosystems in the Age of Agentic Intelligence - Session
32mKeynote Panel Discussion: Three Forces Making Confidential AI A Mandate
19mHow Can We Have Private Access to Intelligence?
40mIn AI We Trust? Securing the Future, One Agent at a Time