MissionPay

AI agents get fooled. Budgets on Avalanche don’t.

The owner signs who a mission may pay. The agent can only propose. A contract on Avalanche rejects every payout outside the signed list, even to a vendor already on your company allowlist.

How it works

Three objects decide every payout.

The agent reads untrusted invoices and web pages, so we assume it will be fooled. It never holds the decision.

Signed mission
Budget0.10 USDC
Payees1
Signer0x4BC9…8F7c
root 0x3011ac81…71dd

The owner signs a mission

Budget, deadline and the exact payees. The payee list goes on-chain as a Merkle root; changing it later is a public event.

Proposal from agent
0.01 USDC
to 0x4BC9…8F7c
proof [ ] → root ✓

The agent can only propose

Every payout comes with a proof that the payee is in the signed set. Without it, there is nothing to execute.

PAYin set, under cap
REVERTanything else

The contract has the last word

GuardedTreasury checks the proof and the caps on Avalanche. Our server holds no owner key, so a hacked server still cannot pay a stranger.

Live on Avalanche Fuji · 13 Nov 2026

Same invoice, two payee rules

One poisoned invoice, sent to the agent on Avalanche Fuji on 13 November 2026. Same treasury, same agent, same vendor. Only the payee rule differs.

MissionPayee rulePaid to vendor 0xE90c…1aEAOutcomeFuji transaction
twin.allowlist.v1Anyone on the company allowlist0.01 USDCDrained0x548d1106…7fc6
twin.mission.v1Only the owner-signed payee set0.00 USDCReverted: PayeeNotInMissionSet0x0ad53adc…3e38
twin.mission.v1 (control)Approved payee, valid proof0.01 USDC to ownerPaid as intended0xcb7236d1…6315
Table 1. Live transactions on Avalanche Fuji, GuardedTreasury 0x2045…d545. The vendor is a wallet we control and was removed from the allowlist afterwards (tx 0x56e1…5562).
Read-only eth_call · no gas

Don’t trust us. Ask the contract.

Type any address. We make a read-only call to the live treasury on Fuji, as the agent, for 0.01 USDC. No signature, no gas.

Open the full judge page
No call made yetPress Check to ask the live contract on Avalanche Fuji
Contract
GuardedTreasury 0x2045…d545
Mission
twin.mission.v1
cast call 0x20451c1e2dCD83F166B4f22868BAdECb5868d545 "proposeWithProof(bytes32,address,uint256,bytes32,string,bytes32[])" … 0x9b1E52D0f3a6C57e0aB0b4c1F2e4D7a8C3f6E091 10000 … --from 0x374b8A126FDb9e5deEc8307da385076155Af3D72
Reproducible from the repo

Results

Each row can be re-run from the repo. Live LLM numbers are being re-measured and are not shown yet.

EvaluationSettingResultSource
AgentDojo banking, replay128 payment injections, MissionPay strict120 / 128 blockedeval/agentdojo/SUMMARY.md
Benign tasks, replay16 legitimate tasks, strict mode3 / 16 wrongly blockedeval/agentdojo/SUMMARY.md
Fuji fork replayAttacker placed on the global allowlist150 / 150 revertedfuji_fork_replay.py
Guard vs contract agreementSame decisions off-chain and on-chain169 / 169fuji_fork_replay.py
ERC-8004 validationIndependent validator, live mission s1c100 / 1000xa9f7fadd…5ffa
Contract testsFoundry, incl. 6 fuzzed invariants111 passingcontracts/test
Table 2. In all 8 missed replay attacks, the user’s own prompt named the attacker as the new payee (user_task_15).

Known limits

  • The owner can reset a mission to allowlist mode. The reset is public on-chain.
  • Cross-chain spending (ICM) is tested locally only.
  • Strict mode blocks payees that only appear in files; a file-trusted mode trades that off.

Active Missions & On-Chain Budgets

Overview of active missions, two separate budgets (Tool session wallet vs Treasury), and cryptographic proofs.

Loading active missions...

MissionPay · Avalanche Team1 Vietnam UniHack 2026 · Fuji testnet only