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.
Vendor invoices
tx 0x0ad53adc…3e38
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.
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.
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.
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.
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.
| Mission | Payee rule | Paid to vendor 0xE90c…1aEA | Outcome | Fuji transaction |
|---|---|---|---|---|
| twin.allowlist.v1 | Anyone on the company allowlist | 0.01 USDC | Drained | 0x548d1106…7fc6 |
| twin.mission.v1 | Only the owner-signed payee set | 0.00 USDC | Reverted: PayeeNotInMissionSet | 0x0ad53adc…3e38 |
| twin.mission.v1 (control) | Approved payee, valid proof | 0.01 USDC to owner | Paid as intended | 0xcb7236d1…6315 |
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- Contract
- GuardedTreasury 0x2045…d545
- Mission
- twin.mission.v1
cast call 0x20451c1e2dCD83F166B4f22868BAdECb5868d545 "proposeWithProof(bytes32,address,uint256,bytes32,string,bytes32[])" … 0x9b1E52D0f3a6C57e0aB0b4c1F2e4D7a8C3f6E091 10000 … --from 0x374b8A126FDb9e5deEc8307da385076155Af3D72
Results
Each row can be re-run from the repo. Live LLM numbers are being re-measured and are not shown yet.
| Evaluation | Setting | Result | Source |
|---|---|---|---|
| AgentDojo banking, replay | 128 payment injections, MissionPay strict | 120 / 128 blocked | eval/agentdojo/SUMMARY.md |
| Benign tasks, replay | 16 legitimate tasks, strict mode | 3 / 16 wrongly blocked | eval/agentdojo/SUMMARY.md |
| Fuji fork replay | Attacker placed on the global allowlist | 150 / 150 reverted | fuji_fork_replay.py |
| Guard vs contract agreement | Same decisions off-chain and on-chain | 169 / 169 | fuji_fork_replay.py |
| ERC-8004 validation | Independent validator, live mission s1c | 100 / 100 | 0xa9f7fadd…5ffa |
| Contract tests | Foundry, incl. 6 fuzzed invariants | 111 passing | contracts/test |
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...