RuntimeGuard-AI: Cryptographic Compliance
summary
RuntimeGuard-AI produces cryptographic verifiable receipts for every consequential agent decision: signed digests of prompt, retrieval context, policy version, tool calls and outputs, stored as Merkle leaves with optional zero-knowledge proofs and external anchoring (transparency log, notarization). Prototype uses a Rust policy engine, emits async evidence off the inference path, and reports latency rising from ~2.3% to ~4.1% plus a 1.4s background proof cost at 50k events; tiered controls (low/medium/high) guide retention and human gates.
AI-written summary · about
chapters (13)
- 0:00Why Logs Aren’t Evidence
- 1:13Runtime Guard Overview
- 1:59Threat Model for Agents
- 4:10Three Layers of Control
- 6:23Agent Identity and Delegation
- 8:26Multi-Agent Trust Boundaries
- 9:48From Docs to Executable Policy
- 11:52Latency and Evidence Path
- 15:25Receipts vs Traditional Logs
- 18:06Privacy, Merkle, and ZK Proofs
- 20:26Risk Tiers and Guardrails
- 22:39Implementation Steps and Anchoring
- 24:04Verify Early and Wrap Up
related talks
28mAgent Runtime Security: Enforce, Attest, Decide
23mGoverning AI Agents at the Hardware Boundary
33mAgentCrypt: Advancing Privacy and (Secure) Computation in AI Agent Collaboration
27mFrom Perimeter to Silicon: Why Agentic AI Demands a New Security Model
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...
24mPanel Discussion: Proof in Production: Customer Lessons from Confidential AI
38mSecurity for AI Agents Using an Ensemble of Fine-tuned... (Lidan Hazout, Bar Kaduri)
17mBreaking the Lethal Trifecta: Secure Agentic Computing via...