AgentCrypt: Advancing Privacy and (Secure) Computation in AI Agent Collaboration
summary
Claims a practical framework called AgentCrypt that enforces 100% privacy guarantee for tagged data by replacing OBO tokens with ID-only auth plus deterministic encrypted responses, adds a three-tier privacy/functionality architecture (retrieve-only level two up to compute-and-encrypt level three), demonstrates implementations with LangGraph and Google ADK, and shows live demos: public-repo push becomes ciphertext and name-similarity salary lookups fail but remain unreadable. Major caveats: per-data policy tags and key management required, and output leakage still needs DP or FHE.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:45AlgoCRYPT research areas
- 2:49AI agent privacy risks
- 4:16LLM privacy scanners limits
- 6:45OBO tokens problems
- 10:56OBO token privacy challenges
- 13:00AgentCrypt framework
- 14:23Architecture with policy engine
- 17:17Data retrieval with encryption
- 22:58Reasoning with private data
- 24:02Private code handling demo
- 27:10Salary data leak prevention demo
- 30:02Credit card leak prevention demo
- 31:09AgentCrypt benefits summary
- 32:12AgentCrypt limitations and future
- 33:00Q&A
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...
26mTrust Is the Next Bottleneck: Why the Agentic Economy Needs Confidential
12mKeynote: Protecting Data in the Agentic AI Era
25mImproving Safety with Privacy: An Exploration of Privacy...
41mAgentic AI, confidential computing and attested TLS
39mTrust Cards For AI
40mIn AI We Trust? Securing the Future, One Agent at a Time