Breaking the Lethal Trifecta: Secure Agentic Computing via...
summary
Claims a practical Causal Isolation Architecture that splits reading, judging, and speaking across isolated components and uses a rule-of-two release judge to block private facts. Open-source repo includes an interactive sandbox, deterministic and real-model evaluation harness, and counterfactual release checks. Evaluation runs ~275 checks, reports 0.0% observed secure-path leakage on claim-eligible OpenAI and Anthropic rows with measured utility 83–100% (model runs take ~14s to 2+ minutes). Built system demonstrably blocks leaks while retaining high utility.
AI-written summary · about
chapters (11)
- 0:00Agentic computing basics
- 1:47Lethal trifecta and prompt injection
- 3:55Incident response and data leaks
- 5:18Rule of two for secure agents
- 6:01Private worker flow steps
- 7:50Counterfactual testing secrets
- 9:56Trust levels and disclosure
- 12:01Open source demo
- 13:29Privacy-utility trade-offs
- 15:15Open source release
- 16:41Conclusion & Q&A
related talks
01h 03mBuilding Secure Infrastructure for Productive AI Agents
34mZero Secrets: Building Secure Infrastructure for AI Agents
27mFrom Perimeter to Silicon: Why Agentic AI Demands a New Security Model
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
12mKeynote: Protecting Data in the Agentic AI Era
33mAgentCrypt: Advancing Privacy and (Secure) Computation in AI Agent Collaboration
48mSecuring Agentic AI
36mSecurity for AI and Multi-Party Collaboration with Confidential Computing (Oh, and...