What the call really does
The calldata is decoded, not trusted. Unlimited ERC-20 and Permit2 approvals, role grants, zero-slippage swaps and reserve manipulation are caught even when they target a contract you allow.
Before your agent pays, GuardianAI decodes what the call really does, checks who it pays and applies your rules for that agent. The agent's wallet won't move funds without GuardianAI's signature for that exact call, so even a fully fooled agent can't pay on its own.
12 contracts live on Monad testnet|271/271 Hardhat tests|2-of-3 Safe owns the contracts|Public drain-the-vault bounty
Architecture note: This sandbox runs the fast-path screen and de-obfuscation in your browser. The semantic classifier, output redaction and the on-chain PolicyGuard check run in the full gateway; try them in the Attack Lab.
funds an agent can move
without GuardianAI's signature
contracts live on Monad testnet
check them live ↗
Hardhat contract tests passing
test snapshot
Safe multisig owns the contracts
since 5 Nov 2026
DISCLOSURE: Compliance and audit artifacts are internally verified empirical test suites (3,211 prompts across 8 corpora, 1,490+ Python & 271 Hardhat tests, 192 Mera Passkey PRF assertions), not third-party statutory certifications.
A payment is signed only if all four pass. The wallet checks that signature on-chain, so the agent can't skip GuardianAI, however it's been tricked.
The calldata is decoded, not trusted. Unlimited ERC-20 and Permit2 approvals, role grants, zero-slippage swaps and reserve manipulation are caught even when they target a contract you allow.
Recipients are checked against GuardianAI's threat feed before anything is signed. The guarded-wallet middleware also flags lookalikes of addresses you trust or have paid recently, a common way to poison an agent's address book.
Per-agent rules in the relay, plus daily outflow caps and a contract and function allowlist enforced by the wallet itself on-chain.
Each agent carries a passkey-signed ID card and an ERC-8004 passport. Strict mode refuses any agent that isn't registered, so an impostor can't borrow a real agent's limits.
The relay decides whether a payment is right and signs it. The wallet refuses anything unsigned or over its limits, so a failure on one side doesn't open the other.
Everything above runs today on Monad testnet. These are next, and we'll mark each one live only when you can try it.
GuardianAI as a policy co-signer for Privy, Turnkey, Coinbase CDP and Safe, so you add the check without moving funds to a new wallet. Base support comes with it.
Each payment is simulated and compared with what the agent was asked to do ("buy API credits from X, up to $20 a day"), not only with static limits.
Payments outside policy go to the owner for a one-tap passkey approval, so a strict policy doesn't stall a legitimate agent.
Scam recipients, malicious payment endpoints and drain patterns seen by one agent are refused for all of them.
Every figure on this site cites its dated repository artifact — and the proof page publishes full benchmark tables, not just highlights.
3,211 prompts across 8 public datasets, August 2026. AdvBench 99.0% strict, JBB PAIR+GCG 98.7% — and Do-Not-Answer's 57.1% published too, because selective reporting is how trust dies.
One external review of the off-chain gateway (March 2026, summary published, zero open critical/high). Internal off-chain, contract and architecture audits July–September 2026, including the Mera passkey enclave audit (166/166 assertions). The Monad contracts have not been externally audited yet.
Twelve contracts live on Monad Testnet, each checked live in the Explorer proof ledger. State anchoring runs simulated by default. No mainnet claims.
The Explorer at aiguardian.dev/explorer runs real attacks against GuardianAI's gates and calls the deployed contracts directly from your browser.
Pick an attack (hidden instructions, a scam invoice, a forged approval, an impostor agent) and watch both gates refuse it: the off-chain firewall first, then PolicyGuard on Monad, with the real call and revert shown.
Launch an attack ↗Every contract's code and live settings, read from Monad testnet each time the page loads, plus a real approved payment.
Open the proof ledger ↗Connect a wallet (read-only), see agent ID cards, check any address against the on-chain scam list, and read the payment guard's live settings.
Open the Console ↗Running agents that hold real funds? Tell us what nearly went wrong. We're choosing the next integrations based on it.