Skip to content
QORECHAIN
Register

Post-Quantum AI Agents

AgentGuard PQ

The AI-agent guardian whose limits live on-chain — enforced by QoreChain's post-quantum smart accounts, not by promises.

Share on X

About

AI agents can hold wallets and move money. When an agent is tricked, injected or simply wrong, who stops the payment — and who can prove what happened afterwards? AgentGuard PQ is a guardian that sits between an AI agent and its money on QoreChain — and unlike a software policy file, its limits are enforced by the chain itself: • The agent's funds sit in a QoreChain abstract account whose spending rules are set on-chain (per-transaction limit 1 QOR, daily limit 2 QOR, QOR only). The guardian's off-chain policy mirrors them, and the dashboard proves the two match by reading the chain keylessly. • The agent never holds the owner's key. It acts through a scoped authenticator key (ed25519, send permission only, 7-day expiry, possession-proof registered) — and the chain itself rejects anything outside the rules. We prove it: a bypass battery of five attacks (over per-tx limit, over daily limit, post-revocation execution, a classical-key-only "Q-Day" send, an EVM send with no authorisation window) was refused by the chain, 5/5. • The EVM lane is bounded by a scoped authorisation window (300 blocks, 5 transactions, 1 QOR cap, native-only) that the owner opens and closes with hybrid-signed messages. • Every agent proposal is signed by the agent's key; every human approval is hybrid-signed (secp256k1 + ML-DSA-87) and both halves are verified. Every verdict — APPROVED, BLOCKED or REVIEW — is written to a hash-chained ledger and anchored on-chain as a hybrid-signed transaction. Run `npm run verify` in the repo: it re-reads the chain keylessly and checks every anchor, the rules mirror and the kill-switch state. No trust required. • The kill switch is on-chain: revoke the authenticator + close the window, and the agent is powerless — demonstrated live, with the chain's refusal recorded. Built with the official @qorechain/sdk on the Diana testnet. Testnet only; all amounts are test QOR with no value. Honesty note (also in the README): two of the five bypass attempts were rejected at broadcast and produced no transaction — the chain's verbatim errors are recorded in data/onchain-proof.json. The chain's AI risk-score precompiles are not yet active on testnet (verified live), so the risk preflight runs on the guardian's own engine and lights up unchanged when they activate.

How it's made

built with the official @qorechain/sdk on the Diana testnet; funds in a QoreChain abstract account with on-chain spending rules (1 QOR per transaction, 2 QOR daily, QOR only); scoped ed25519 authenticator key for the agent; hybrid secp256k1 + ML-DSA-87 approvals; hash-chained verdict ledger anchored on-chain; scoped 300-block EVM authorisation window; kill switch by revoking the authenticator and closing the window; verified with `npm run verify`.

Timeline

Comments and questions

  • No comments yet.