Read what a proposal does before anyone votes on it.
DAOs are starting to let AI read and vote on proposals, and the proposer writes the text the AI reads. QuorumGuard decodes the on-chain action, compares it with the description, and flags hidden instructions, lookalike accounts and amounts that do not match.
Every new NEAR DAO proposal, checked as it lands.
QuorumGuard follows NEAR mainnet block by block. When anyone opens a proposal in any of the 2,098 Sputnik DAOs, it is decoded and checked within seconds, long before the vote closes. Councils can get these alerts in Telegram.
31 poisoned proposals, 15 honest ones.
Seven attack families against an AI that reads proposals: hidden Unicode, hidden markup, fake authority, says-one-thing-does-another, high-power actions dressed as routine, other languages, and plausible cover stories. The honest set copies how real DAO members write: links, dollar conversions, "400N", emojis, text pasted from X.
Real proposals, real DAOs, all approved.
We ran QuorumGuard over every proposal of three busy mainnet DAOs. Most pass. The ones it stops are worth a question before a vote, and an AI voter would have waved them through. These are questions, not accusations: most are probably typos. That is the point: a typo in a receiver sends real money to someone else.
Open source (MIT). Read-only: QuorumGuard never signs anything for you. Verdicts: BLOCK means an AI must not vote on it alone; WARN means a human should read it; CLEAN means the text matches the action.