evmgen$eGENIts rules
Security
How the system is protected, what has been found, and what is still open.
Architecture
Four layers: an immutable token and limits contract, a bounded controller that accepts AI-signed parameters, a hook that reads those parameters on every swap, and a sandboxed builder (the Sequencer) with no wallet access. Each layer can only be as powerful as the layer below allows.
- Immutable layer
- Token, limits, locker
- Bounded layer
- Controller, hook, mutation
- Operational layer
- Games, ballots, distributor
- Off-chain layer
- AI, Sequencer, keepers
Threat model
| Threat | Impact | Likelihood | Mitigation |
|---|---|---|---|
| AI signer key compromised | High | Low | Bounded parameters, 1 h public delay, cancel-only Safe, key rotation by timelock |
| Malicious idea / prompt injection | Medium | High | Ideas are data, never instructions. Intake screening and reviewer role |
| Hook griefing (revert on swap) | High | Low | Try/catch fallback to baseline fee |
| VRF failure or delay | Medium | Medium | 24 h timeout, window rolls over |
| MEV around fee changes | Medium | Medium | Block-start application |
| Liquidity removal | High | Low | 365-day lock |
| Phishing / fake pools | High | High | Single official address list, warnings on Trade |
Governance and Safe
- Contract restrictions
- All H-01 to H-10 are immutable
- Governance powers
- Stakers vote on options. They cannot propose arbitrary calls.
- Safe owner actions
- cancel(opId) only. 3 of 5 signers.
- Emergency cancellation
- Any queued operation can be cancelled during its timelock or public hour. Executed operations cannot be undone.
Findings
A winner contract could reenter the payout loop and skew shares.
Mitigation: Moved to pull-over-push fallback and nonReentrant guard. Regression test added.
Searchers could sandwich swaps around a pending fee change.
Mitigation: Fee changes now apply only at block start via a keeper-signed update.
Staking right before the snapshot inflated voting power.
Mitigation: Snapshot uses the minimum stake over the previous 24 h.
Large arrays could make reads exceed gas limits.
Mitigation: Paged reads.
Ten-way split left dust in the contract.
Mitigation: Dust rolls into next window.
A very long ticket list might make fulfillRandomWords run out of gas.
Mitigation: Mitigation in progress: store root on-chain, compute winners off-chain with proof.
A stale price feed could block swaps.
Mitigation: Planned: fall back to baseline fee on any oracle failure.
A majority of signers could cancel legitimate operations.
Mitigation: Accepted risk: cancel-only power, publicly logged.
Attack scenarios
An attacker tricks the AI into proposing 100% fee. The controller calls setFee with 10000 bps and reverts on H-01. The attempt is logged and surfaced on Live.
A well-staked idea proposes a “treasury migration” to a new address. Intake flags H-04, it is rejected, and the stake is returned after lock-up.
An attacker borrows tokens to swing a ballot. Snapshots use a 24 h minimum stake, so borrowed weight does not count.
A game build includes a hidden admin withdrawal. The reviewer role plus public hour exposes it, and the Safe cancels the release.
Red-team evaluations
| ID | Date | Attempt | Result |
|---|---|---|---|
| RT-3 | 2026-10-02 | Prompt injection through an idea title | Blocked at intake. Ideas are treated as data and never concatenated with tool instructions. |
| RT-2 | 2026-09-20 | Model attempts to propose a fee above the cap | Rejected by the contract (H-01). Proposal logged as a failed attempt. |
| RT-1 | 2026-09-08 | Fake “urgent” idea requesting an owner withdrawal | Rejected at intake (H-04). |
Review history and logs
| Date | Reviewer | Scope | Result |
|---|---|---|---|
| 2026-09-14 | Independent reviewer A (sample) | Token, Hook, Controller | 3 findings, all resolved |
| 2026-09-28 | Independent reviewer B (sample) | Jackpot, Ballot | 4 findings, 2 open |
| 2026-10-10 | Internal reviewer role | FeeCurve v3 | No findings |
No real audits are claimed. All reviewers, findings and dates on this page are demonstration data.