Compound is a crypto lending protocol governed by holders who delegate their $COMP tokens, an arrangement known as a decentralized autonomous organization, or DAO. It functions like an online republic: token holders debate proposals, vote, and software carries out the result. DAOs are not a niche experiment; collectively they govern treasuries worth billions of dollars and control some of the largest DeFi protocols in operation, which is why their governance mechanics have real financial stakes attached.

In July 2024, that republic nearly handed a fortune to a small group of voters. Proposal 289 asked Compound to transfer 499,000 $COMP — then worth about $24 million — into a yield-bearing vehicle that those voters controlled. Two earlier versions had failed, and the third appeared headed for the same fate.

Then, in the final 34 minutes, supporting addresses cast 563,591 votes, equal to 82% of all support for the proposal. The last large block arrived eight minutes before the deadline, and the measure passed 682,191 to 633,636.

The episode was extremely controversial and remains highly contested, yet there was no malfunction in the code — it worked exactly as intended. That was precisely the problem: the wallets had accumulated enough $COMP and delegated their voting power before the window closed, and Compound had no emergency authority that could pause the software. Several individually reasonable rules had combined into a convenient path for a treasury raid.

Compound reached a settlement that canceled the allocation and later added a veto role, installing a brake inside a system built on automatic token-holder rule.

That episode captures the central DAO dilemma: most defenses against rushed or hostile votes end up giving somebody more control over participation or the final outcome.

Two 2026 studies — one from the Max Planck Institute for Software Systems and one from Vrije Universiteit Amsterdam — traced a similar problem across 48 large Ethereum DAOs. One examined how registration, staking, and delegation concentrate voting power; the other mapped attacks that rely on valid governance rules.

The ballot has a velvet rope

Calling a governance token a "vote" is not quite accurate. Depending on the DAO, a holder may need to register a wallet, lock tokens, delegate them, maintain a minimum balance, or pay for an on-chain transaction before actually casting a vote.

Proposals face hurdles of their own. Someone needs enough tokens or delegated support simply to introduce one, and the idea may pass through a forum and an informal poll before a binding vote on the blockchain or through an off-chain service such as Snapshot, which many Ethereum DAOs use to poll holders without charging transaction fees.

Once the tally clears the quorum and approval formula, a smart contract, a multisignature wallet, or a named person carries the result into effect.

Each of these gates solves a real problem, but each also favors a particular type of participant. Proposal thresholds discourage spam and malicious code, yet they inadvertently reserve authorship for wealthy holders and established delegates. On-chain voting makes results enforceable, but transaction fees favor those with enough money and conviction to use it. Free off-chain polls attract a wider crowd, then depend on a smaller group for execution.

The researchers found an even split: 24 DAOs used on-chain voting and 24 used off-chain systems.

Uniswap illustrated how different electorates can form inside the same organization: more wallets joined its free off-chain polls, while much larger blocks of voting power appeared during the paid on-chain phase that could make a proposal binding.

Turnout is only a small part of the picture. A protocol may have thousands of token holders while a handful of addresses control proposals, votes, and execution. By the time the public tally appears, the rules have already selected the electorate.

The security rules pick the ruling class

DAOs often keep tokens in treasury contracts, and founding teams or investors may hold allocations that have yet to vest, so registration separates circulating tokens from balances that currently carry voting rights.

Of the 48 DAOs studied, 36 required some form of registration, and only four had registered more than half of their outstanding supply. Across those 36 organizations, the average registered share was 21%, meaning the practical electorate usually covered a small fraction of all tokens.

Much of the missing supply belonged to users whose coins sat on exchanges or in DeFi protocols. Centralized exchanges held more than 10% of outstanding tokens on average across the sample, and DeFi contracts held another 3.5%. In 14 registration-based DAOs, those intermediary wallets controlled more tokens than the entire registered electorate.

This creates a strange and rather unique custody problem: an exchange wallet can represent thousands of customers even though the blockchain sees a single address with one giant balance. Letting the exchange vote turns a custodian into a political heavyweight; excluding it strips customers of governance rights attached to tokens they paid for. Most DAOs also allow one wallet to send all of its power to a single delegate, which makes splitting votes among the underlying owners difficult.

Staking addresses a different vulnerability by making voting power expensive to build and slow to unwind. An attacker can buy or borrow a large position, approve a favorable proposal, and sell once the vote closes; a lock keeps that voter financially exposed to the outcome for longer.

Fifteen DAOs required staking, with a median of 27.4% of tokens locked. Some imposed a one- or two-week withdrawal wait, while Curve, Angle, and Frax offered stronger voting power for locks lasting up to four years. The system rewards patience and turns liquid wealth into a prerequisite for political influence. Curve's model in particular spawned a secondary market in which protocols competed for locked voting power — a dynamic widely known as the "Curve wars" — illustrating how vote-boosting rules can turn governance influence into a tradable commodity.

Crypto soon produced middlemen for people who wanted both influence and the freedom to trade. These services maintain long locks, issue tradable substitutes, and keep the original voting rights. According to the researchers' measurements, the arrangement concentrated enormous voting blocs inside a few services.

Delegation works similarly, because most holders have limited appetite for forum arguments about collateral ratios. Handing votes to a professional participant makes sense, and repeated delegation builds durable political blocs.

The ten largest holders controlled more than half of voting power in 39 of the 48 DAOs, and delegated voting was usually more concentrated than direct voting.

Registration protects treasury balances, staking makes quick attacks costlier, and delegation gives passive holders a voice through someone who pays attention. Put together, the people with the most capital, time, technical fluency, or control over customer assets tend to run the place.

A legal DAO vote can still be a raid

The second paper defines a governance attack as an actor using the authorized process to win an outcome that harms the wider organization.

Across 28 DAO incidents, researchers classified 16 as attacks that a different mechanism could have prevented. Six involved contract bugs, while ten depended on buying or borrowing enough tokens to influence a vote.

This pattern has precedent. In 2022, the credit protocol Beanstalk lost roughly $182 million after an attacker used a flash loan to acquire enough voting power to pass a malicious proposal through the protocol's own governance process — an earlier demonstration that a formally legitimate vote can function as a theft mechanism.

Compound is the clearest example in the study. The wallets associated with Proposal 289 gathered more than 680,000 $COMP over four months. Researchers traced 563,790 tokens acquired through four centralized exchanges and another 118,089 borrowed through Compound itself — even though those addresses had held only 853 $COMP before the buildup and had little history in the protocol's politics.

The late burst exploited a community that expected the third proposal to fail. Compound could have extended the vote when a large bloc appeared near the deadline, required longer staking, or allowed a trusted council to pause execution. Every one of those options would have shifted power toward reactive voters, committed holders, locking services, or a small emergency body. That last option — a security council or guardian multisig with the power to pause or veto — is already a common design in Ethereum ecosystems, and its trade-off is exactly the one Compound faced: it reintroduces a trusted authority into a system designed to eliminate one.

Compound chose the emergency brake. And in the 2024 configurations the researchers reviewed, seven other DAOs shared its exposure to readily available voting power and late vote accumulation: Uniswap, Radicle, Gitcoin, Silo, Ampleforth, Hop, and Cryptex. Those systems can evolve through governance, so the list records a moment in 2024; a current security rating would require a fresh review.

Decentralization needs a richer accounting than token distribution alone. A good governance report would show how much supply can actually vote, how much power the largest delegates control, which intermediaries hold staked tokens, and who can introduce, execute, or veto proposals.

Smart contract audits already ask whether governance code follows its specification; a constitutional audit would ask where that specification sends authority.

DAOs can spread ownership across thousands of wallets and still funnel practical control toward a few dozen professionals, custodians, and large holders — with software that performs flawlessly all the way through.