Leash.

Threat model

Leash protects the funds an AI agent controls. This page lists who can do what, what we assume, and what stops each attack. Every "stopped by" points at a test or a run you can reproduce.

Assets

  1. The guard's funds (native QOR and ERC-20s held by a LeashGuard or LeashGuardV2).
  2. The mandate: the caps, recipients, routes and rate floors the owner signed.
  3. The kill switch: the ability to stop spending quickly.

Actors and keys

Actor Holds May
Agent a secp256k1 key spend and act inside the mandate; freeze
Owner, classical a secp256k1 key (the guard's immutable owner) freeze only
Owner, post-quantum an ML-DSA-87 key registered on chain sign a new mandate (start, loosen, re-arm)
Anyone nothing submit a signed mandate

Assumptions

Attacks and what stops them

# Attack Stopped by Evidence
1 Agent key leaks; attacker spends more than the cap per-action and daily caps, shared by payments and actions test_capsApplyToActions_andShareTheBucketWithSpend, fuzz test testFuzz_dailyCapHoldsAcrossActions
2 Attacker pays an address that is not allowed recipient allowlist drill row RecipientNotAllowed
3 Attacker loosens its own limits applyMandate needs an ML-DSA-87 signature over the whole mandate test_agentCannotLoosenRoutes, drill rows BadPqcSignature (testnet: real chain)
4 Owner's classical key leaks; attacker grants itself power classical key can only freeze; applyMandate has no classical path drill: "owner key grant itself a bigger mandate" is blocked
5 Replay an old or foreign signature the signed message binds the chain id, this guard, a nonce and the mandate hash test_mandateNeedsTheOwnersPqSignature; drill: replay from a different guard blocked
6 Quantum computer breaks secp256k1 both classical keys can only stop things; loosening is post-quantum design; the whole drill
7 Router under-delivers or lies about a quote rate floor checked on the measured balance change test_rateFloorIsEnforcedOnWhatWasReceived, test_vaultThatSkimsDepositsIsCaught; testnet drill row BelowRateFloor
8 Router sends the output to the attacker output recipient must be the guard, and received amount is measured test_outputMustGoToTheGuard, test_routerThatDivertsOutputIsCaughtByMeasurement
9 Call data that the adapter does not understand adapter reverts; no route, no action test_unsupportedCallDataIsRefused; drill row BadAction
10 Action through an unlisted router or adapter route lookup test_unknownRouterOrAdapterIsNotARoute; drill row RouteNotAllowed
11 Approval left open to a router approvals are exact and reset to zero after every call test_noApprovalIsLeftOpen_evenIfTheRouterPullsLess
12 Reentrancy through a malicious target lock on spend and execute; callers other than the agent are refused anyway test_reentrancyIsBlocked
13 Target is the guard itself or a precompile refused test_cannotTargetTheGuardOrPrecompiles
14 Stale mandate keeps spending expiry; the guard starts frozen test_expiryStopsActions
15 Fast drain by a leaked key before anyone notices the monitor flags bursts and unexplained outflows and can freeze with a stop-only key pnpm demo:monitor; tools/monitor.test.ts

Not stopped, on purpose or by limitation

Limit Why it matters What to do
No price oracle. A rate floor is the owner's own number. A stale floor blocks good trades or allows bad ones. Market moves and sandwiches inside the floor are not prevented. Set tight floors, short expiries, small caps; review floors when prices move.
Route targets are trusted. A malicious or upgraded router is bounded by the floor and the caps, not removed. Route only to contracts you have read; prefer immutable routers.
Owner post-quantum key leak. The attacker can sign any mandate. Hardware-backed storage; short mandates; the monitor flags mandate-changed; freeze first, then rotate the guard.
Owner post-quantum key loss. Mandates cannot change. The kill switch still works. Deploy a new guard and migrate.
Funds outside the guard. Limits cover only what the guard holds. Keep the agent's own account near zero.
Fixed 24-hour daily bucket. An agent can spend one cap at the end of a bucket and another at the start of the next. Choose caps that are safe doubled.
Window renewals. The chain caps an EVM window at about 6 hours; long mandates need a keeper holding the agent account's ML-DSA-87 key. Keep that key off the agent's machine.
Unaudited code. Bugs may exist despite 46 tests and mutation checks. Independent audit before any mainnet use (AUDIT_PACKAGE.md).
Testnet only. pqcVerify is not active on mainnet. See ROADMAP.md.
Native AMM and bridges QoreChain's native AMM cannot be called from a contract; no bridge route is live. Out of scope. Leash does not claim to guard them.

Defence in depth inside execute

After every action the guard checks that no other listed token's balance fell and that the spent token fell by at most the declared amount. With well-behaved tokens a route cannot trigger these checks, so our tests cover them only as far as the mocks allow. They are there for the day a listed token or target misbehaves.

View this page on GitHub · raw markdown