Chain of Custody — Cryptographic Proofs for AI Agents (elsium-ai)

Every action your AI agent takes — cryptographically signed, hash-chained, and verifiable offline. Capability tokens scope what it CAN do; signed proofs record what it DID. A live demo of the trust layer most agent stacks are missing.

In this video I walk through a complete demo built on elsium-ai 0.15 that closes the loop between authorization and audit for LLM agents:

▸ Ed25519-signed capability tokens — scope tools, models, RAG stores, data classes, cost & TTL
▸ Delegation — mint child tokens with strictly tighter scope from a parent
▸ Revocation — kill a token mid-flight; subsequent runs are rejected before any tool fires
▸ Verifiable Execution Proofs — every LLM call, tool call, RAG retrieval & policy decision goes into a SHA-256 hash chain, signed by the org key
▸ Offline verification — auditors verify the entire run with only the public key, no API keys, no network
▸ Tamper detection — flip a single byte and `verifyProof()` rejects at the exact broken event index

⏱ Chapters
00:00 — Why agents need a trust layer
00:30 — CH.01 Happy path: an agent does its job, signed proof recorded
01:00 — CH.02 An attacker flips a byte → the auditor catches it offline
01:35 — CH.03 Same token, forbidden field — the denial is recorded too
02:00 — CH.04 Delegating a stricter child token to a sub-agent
02:35 — CH.05 Revoke. Effective immediately.
03:00 — Recap: three primitives, one trust loop

🛠 Built with
• elsium-ai 0.15 — open-source TypeScript framework for production AI agents
• @elsium-ai/core — Ed25519 signer, KeyRegistry, CapabilityIssuer + Verifier, RevocationStore
• @elsium-ai/observe — ProofRecorder, verifyProof, compareProofs
• @elsium-ai/tools — withCapability tool guard
• Hono + Bun — server-sent events for the live proof stream

🔗 Links
• elsium-ai on GitHub: https://github.com/elsium-ai/elsium-ai
• Capability Tokens reference: https://github.com/elsium-ai/elsium-ai/tree/main/examples/capability-tokens
• Verifiable Execution reference: https://github.com/elsium-ai/elsium-ai/tree/main/examples/verifiable-agent-execution
• Other demos in this channel: [add your playlist URL]

Who this is for: anyone shipping LLM agents into regulated workflows — fintech, healthcare, legal, public sector — where "the model said so" is not an audit answer. Also: red-teamers, AI infra engineers, and anyone tired of agents that can call any tool with any args.

Видео Chain of Custody — Cryptographic Proofs for AI Agents (elsium-ai) канала Futurmille AI Studio