Bribery in DAOs

Introduction

Standing assumption. At least one eligible voter casts a positive-weight yes or no ballot and neither makes nor accepts a vote-contingent offer. This applies to every modeled election, including the no-offer baseline. In comparisons that hold some ballots fixed, we call these the background ballots. All models use weighted majority without a quorum: only the weights of the yes and no ballots determine the outcome. When identifying a historical outcome, we use the relevant governor’s rules or recorded state. Where we compare only scores or majority margins, we say so.

Token-weighted governance allocates voting power by stake and is common in blockchains. In practice, participation can be low, and large holders can exercise substantial control [1]. As a result of nonvoting and concentration, shared assets can be left vulnerable to governance attacks. In July 2026, BonkDAO reported that a malicious governance proposal had drained roughly $20 million of BONK from its treasury [2].

Abstention by small holders can be individually rational. Voting is implicitly costly, as it requires learning about and casting the vote. Additionally, a vote cast by a small stakeholder is unlikely to be pivotal (i.e. to change the result) when large holders or a few blocs participate [3], [4]. Concentration thus discourages participation from smaller stakeholders. Following VBE [5], we call the aggregate stake of abstaining holders the inactivity whale.

Under bribery, an actor compensates voters for participating, abstaining, or voting a particular way. In traditional elections, secret ballots make votes hard to verify, so bribers tend to mobilize inactive supporters with similar preferences [6]. Such exchanges can benefit both briber and voter, but worsen collective outcomes and can be associated with lower electoral trust [7], [8].

On DAOs however, voting may be more transparent. Public voting records permit bribes conditional on verifiable choices, and Dark DAO mechanisms can enforce bribes while concealing bribery [5]. V. Buterin [9] argues that coin voting is especially exposed: a voter’s chance of being decisive is tiny, so a bribe can exceed her expected loss from a bad decision, even when the harm to everyone else far exceeds the bribes paid. LobbyFi, for example, sells the use of delegated Arbitrum voting power for one proposal. In one reported purchase, the buyer paid about $10,000 to direct votes associated with approximately $6.5 million of ARB [10]. The tokens themselves did not change hands in that purchase.

The effect of such bribery on governance is contested. Critics, following Buterin, warn that it detaches voting power from exposure to outcomes, while proponents argue that it activates passive holders and raises turnout [11]. Observed purchases establish that voting rights can be bought, but does not establish when a single buyer can profitably mobilize inactive holders or whether the resulting elections better reflect participants’ preferences. Because pivotality depends on everyone’s strategies, answering these questions requires an equilibrium analysis.

We ask whether allowing a single actor to pay holders to vote for a proposal increases Nash social welfare. We begin with one buyer and one binary, stake-weighted decision. Each voter has a private value drawn from their own distribution. The buyer and a recipient may benefit from a transaction while the resulting decision harms other holders, so the welfare question depends on the external effects of the vote. With fixed unpaid background ballots, we show that the Nash welfare of the voters who can receive offers cannot fall when those voters would otherwise abstain and each has a nonnegative prior mean value for passage. Offers below the voting cost give a strict increase if they induce a new ballot and values are atomless. These conditions allow realized disagreement; they do not cover every governance attack.

Models of costly turnout with private information are standard in voting theory [12]. A related vote-buying model studies a single buyer and voters with privately drawn preferences, but does not model abstention [13]. Some computational bribery work instead asks how hard it is to change an election outcome given votes and prices [14]; we do not study computational complexity here.

We also ask whether decentralized payments can defend against malicious proposals when the attacker can buy votes too. This is a second model: agents can make additive, direction-contingent offers to any eligible voter. Its results concern blocking capacity and funding incentives. We also ask when paid turnout complements active delegates, and when the resulting transfers make holdings more equal. A profitable intervention alone does not establish equilibrium protection.

Historical cases and observed markets

These cases motivate the model and show where its assumptions can fail. They are selected illustrations. The panel uses its own inclusion rule; it does not add a proposal because it is discussed here.

Curve, Convex, and Frax. Vote incentives can be paid by a single protocol without purchasing holders’ tokens. In an analysis of Curve, Convex, and Votium data through March 17, 2023, Frax was the largest contributor of incentives to vlCVX holders, accounting for approximately $103.69 million, or 41% of those incentives in the study. In the mature part of the sample, a gauge’s share of incentives and its share of votes had a correlation of 0.99 [15]. This evidence covers many buyers, gauges, and voting rounds; it does not identify a single attack or establish that the incentives caused the votes. Frax’s activity is a documented vote-buying case, not a demonstrated malicious governance attack.

The same study reports 1,124 unique addresses participating in Convex gauge proposals and an average of 77 addresses in non-gauge proposals [15]. These are different proposal types, with emissions and other incentives as well as bribes; the comparison does not estimate the turnout effect of one buyer. Addresses need not correspond to distinct voters. Curve gauge votes are also multi-option and mediated through Convex, unlike our binary, single-buyer model.

Compound / Golden Boys.

Proposal 289 sought 499,000 COMP and was reported as passing with approximately 682,191 for and 633,636 against [16], [17]. Holding that tally fixed, roughly 48,555 additional no units would have tied it. We replayed the governor ballots and COMP checkpoints. Of 58 ballot submissions, 44 had positive weight: 23 yes, 20 no, and one abstain. The largest nonvoting snapshot delegate alone had 256,007 COMP in voting power, more than the 48,554.70 COMP needed to tie. Of 2.42 million eligible COMP, 1.10 million went uncast. The proposal was later canceled following negotiation [18]. The votes do not establish which accounts had common control, whether anyone was paid, or whether a nonvoter was recruitable.

Figure 1: Compound Proposal 289: reported final voting weights [16], [17]. The approximate gap is 48,555 COMP. This is a retrospective tally comparison, not evidence that enough additional opposition could have been recruited before the deadline. The proposal was later canceled.

Indexed Finance.

On 16 November 2023 a single address cast the only yes ballot on Proposal 24; voting closed on 18 November. A published case study links that vote to an attempt to seize timelock control and describes the appeal to voters. The contract records 402,219 NDX for and 413,036 against. Of 337 eligible addresses, 11 cast positive-weight votes (one yes and ten no); another address submitted a zero-weight no ballot. The first no vote landed 6 hours 28 minutes before close. A documented public appeal came 3 hours 30 minutes before close, so some opposition was already active by that appeal. The two largest no votes supplied about 95% of all opposition; removing either while holding the rest fixed would have reversed the tally. This is evidence that late mobilization can defeat an attack. We found no verified payment for those no votes.

The second attack, Proposal 27, finished 24,989 NDX ahead. The largest of its nonvoting delegates had 100,444 NDX in snapshot weight and total uncast weight was 603,026 NDX. The earlier community proposal to disable timelock control (26) had already cleared the vote and could execute first. Proposal 27 was canceled; the published case study reports that the attacker separately received about $10,000 through escrow for withdrawing. This is a payment to withdraw a proposal, outside the ballot contracts in our model.

Actual vote purchases.

The Arbitrum Foundation reports a LobbyFi purchase of 19.3 million ARB voting units for 5 ETH, and another purchase of about 20.1 million units for 0.0652 ETH [10]. This establishes priced voting power, not a turnout elasticity or defensive success. The advertised pooled winner-take-all auction with withdrawable losing bids also differs from our additive bilateral contracts. Balancer’s Humpy treaty explicitly discusses vote incentives [19], but concerns repeated gauge allocation rather than one malicious binary decision. Neither case establishes that payments generally go from richer owners to poorer owners.

Figure 2: Two Arbitrum/LobbyFi purchases reported by the Foundation [10]: 0.0652 ETH for approximately 20.1 million voting units and 5 ETH for approximately 19.3 million. The normalized prices differ by about 80 times across different decisions. These are observed transaction prices, not causal turnout responses or competing bids in the same auction.

Limits of the attack class.

Beanstalk’s April 2022 disclosure identifies a flash-loan governance exploit [20]; it does not validate a fixed-weight protective-voting window. Mandiant reports 1.2 million fabricated Tornado Cash votes against approximately 700,000 legitimate votes [21]; after that capture, full legitimate turnout would not suffice under a simple-majority comparison. Build Finance and Synthetify motivate proposal detection and attention, based here on secondary reporting [22], [23]; their historical capacity and timing have not been replayed.

BONK. Reports on BIP 76 put the tally at 882.38 billion BONK for a quorum of 879.95 billion and the transfer at 4.43 trillion BONK. The proposal was filed on 30 June 2026, and the decisive stake was reportedly assembled on 4–5 July, before the 6 July vote and execution [2] CoinDesk. These reports describe token acquisition during the proposal’s life and a promised yes-voter reward, but do not verify a paid ballot or a fixed eligible opposition set. We treat BONK as a boundary case for the snapshot assumption; we do not use it in the fixed-panel defense replay.

Figure 3: Tornado Cash: reported governance power after votes were fabricated [21]. These are not cast-vote totals from the original proposal. The comparison illustrates a capacity failure after capture, not the earlier opportunity to detect and reject the malicious payload.

Single-buyer model and welfare

The first model retains one buyer and a common per-unit offer. The welfare comparisons allow active voters at baseline. For the Nash-welfare benchmark, the voters who can receive offers start out inactive, while the unpaid background ballots remain in the tally. We study the distribution of holdings separately.

Model

There are 𝑛 voters who can receive offers, in addition to the voters casting fixed background ballots and the buyer. There is one proposal, with outcome 𝑥=1 if it passes and 𝑥=0 otherwise. Voter 𝑖 has public stake 𝑠𝑖>0 and privately observes 𝑢𝑖, their utility per unit of stake for passage relative to rejection. Values are drawn independently, with 𝑢𝑖∼𝐹𝑖; the distributions may differ across voters and are common knowledge.

Each voter votes yes, votes no, or abstains. Voting costs 𝑐𝑠𝑖, where 𝑐>0 is a common cost per unit of stake; abstention is free. The side with greater total voting stake wins, and a fair coin resolves ties in this benchmark.

An eligible voter is the only buyer and values passage at 𝑊>0. Its own ballot is fixed in this benchmark and counted separately. The fixed background voters make and receive no offers. The buyer publicly offers 𝑏𝑠𝑖 to voter 𝑖 for voting yes, where 𝑏≥0 is a common payment per unit of stake. The payment is enforceable and does not depend on whether the proposal passes. Voters observe the offer and choose simultaneously. They may refuse it and vote freely; if 𝑏>0, accepting is better than casting an unpaid yes vote.

Let ℎ𝑖>0 be voter 𝑖’s utility from rejection before voting costs and payments, and let ℎ𝐵 be the buyer’s baseline utility net of the cost of its fixed ballot, if any. The realized payoffs are

𝑈𝑖=ℎ𝑖+𝑠𝑖𝑢𝑖𝑥−𝑐𝑠𝑖𝟏{𝑖 votes}+𝑏𝑠𝑖𝟏{𝑖 votes yes},𝑈𝐵=ℎ𝐵+𝑊𝑥−𝑏∑𝑖𝑠𝑖𝟏{𝑖 votes yes}.

Voters maximize expected utility given their private values and the other voters’ strategies. The buyer chooses 𝑏 to maximize expected utility. We compare the voting game without an offer to a Bayesian Nash equilibrium after an offer. If the voting game has multiple equilibria, the comparison depends on which equilibria are selected.

Welfare with active voters

Compare two specified continuations of the same proposal, with and without payments. The no-payment continuation may already contain active voters. Let 𝑋0,𝑋1∈{0,1} indicate passage, and let 𝐻 be the net harm from passage to the accounts included in the welfare calculation. Count both sides of each payment so transfers cancel. Let 𝐾0 and 𝐾1 be the real costs of voting and running the mechanism.

Proposition (Welfare test). Expected total welfare strictly increases if and only if

𝔼[𝐻(𝑋0−𝑋1)]>𝔼[𝐾1−𝐾0].

This follows by subtracting the two payoffs after summing across accounts. The left side counts the harm avoided and allows it to vary with the state. The right side counts extra real costs, including any costs saved if an active voter stops voting. If the welfare measure includes the attacker, its benefit from passage reduces 𝐻. An expected welfare gain, a profitable intervention for the payer, and greater turnout are three separate claims.

When recruitment complements an active voter

For a one-sided defense, use weighted majority with ties rejecting. A holder has voting weight 𝑤, incurs cost 𝑘 to vote, and loses 𝐿 if the proposal passes. Let 𝐹(𝑚)=Pr(𝑀≤𝑚) describe the yes lead 𝑀 before adding this holder’s ballot and any recruited weight. Compare against the same distribution of other ballots. Suppose buying 𝑟 additional no weight costs 𝐶 and guarantees those ballots, with no counteroffer in this comparison.

Proposition (Retention and funding). After buying 𝑟 no weight, the holder strictly prefers to vote no if and only if

𝐿(𝐹(𝑟+𝑤)−𝐹(𝑟))>𝑘.

Allowing the holder to choose its own ballot in both cases, define

𝑉(𝑟)=max{𝐿𝐹(𝑟),𝐿𝐹(𝑟+𝑤)−𝑘}.

The holder strictly gains from financing the recruits if and only if 𝑉(𝑟)−𝑉(0)>𝐶. At equality it is indifferent.

Proof. Abstaining gives rejection probability 𝐹(𝑟); voting no gives 𝐹(𝑟+𝑤) at cost 𝑘. Add the common constant 𝐿 to put these payoffs in terms of avoided loss, take the better action, and subtract the cost of the intervention. □

The change in the holder’s pivotality is

Pr(𝑤<𝑀≤𝑤+𝑟)−Pr(0<𝑀≤𝑟).

Recruited votes can make the holder decisive in states where its vote previously was insufficient. They can also make its ballot redundant in states where it previously mattered. If the holder already votes, its pivotality stays strictly above its turnout threshold for every affordable coalition, and some coalition is profitable, then its optimal intervention recruits new turnout and retains its own vote, provided a best intervention exists. With finitely many coalitions at known prices the optimum is attained.

For example, let 𝑀 be 1 or 3 with equal probability, 𝑤=2, 𝐿=3, and 𝑘=1. One neutral voter with weight 1 accepts a payment of 1 to cover its voting cost. The holder votes no both before and after buying that vote: its pivotality stays at one half, but shifts from the low-margin state to the high-margin state. Rejection rises from one half to certainty. The holder gains 32−1=12 after paying for the new ballot.

Nash welfare as a benchmark

The benchmark measures the geometric mean of the positive ex ante utilities of the voters who can receive offers. It does not include the buyer or the voters casting fixed background ballots. Assume |𝑢𝑖|<𝑐 almost surely and 𝔼[𝑢𝑖]≥0 for every voter who can receive an offer. Without offers, those voters all abstain. Let 𝑝0 be the passage probability from the fixed ballots alone: it is zero for a no lead, one for a yes lead, and one half for a tie. With a yes offer, voting no is still strictly dominated by abstention for those voters. By abstaining, any such voter therefore faces passage probability at least 𝑝0, independent of its type. Because its prior mean value is nonnegative, this option gives at least its no-offer ex ante utility. A best response weakly improves on that option, so the Nash product cannot fall. If 0<𝑏<𝑐, values are atomless, and the offer induces a new ballot with positive probability, some voter strictly gains: a yes ballot must have a positive chance of affecting the result to cover its net cost, and the type exactly indifferent between yes and abstention then has probability zero.

The conclusion does not extend to electorates with active voters who can receive offers. For this example, fix two equal-weight unpaid background ballots, one yes and one no. Among the voters who can receive offers, take stakes 2 and 1, unit baselines, and cost 1 per unit. The whale’s per-unit value is 20 with probability 90% and zero otherwise; the small holder’s per-unit value is 1.5. Without an offer, only the high-type whale adds a ballot. With an offer of 0.5 per unit, there is an equilibrium where only the small holder adds a yes ballot: it gains 0.75 in expected outcome value at a net voting cost of 0.5, while the whale can abstain without changing the outcome. Passage rises from 95% to certainty, so the offer is profitable for a buyer with value 20. Yet voters’ ex ante utility product falls from 35.2×2.425=85.36 to 37×2=74. The example compares specified voting equilibria and a profitable offer; it does not require that offer to be optimal.

A fixed panel for the turnout conditions

For each proposal, let 𝑤𝑖 be a delegate’s eligible weight and 𝐼𝑖 indicate whether it voted. Over the same eligible delegate population, the fraction of weight cast minus the fraction of delegates voting is exactly

Cov(𝑤𝑖,𝐼𝑖)𝔼𝑖[𝑤𝑖].

The gap is positive if and only if there is a positive turnout-weight association. This says more than that a few accounts hold much of the weight. It does not identify a causal effect of weight or imply a monotone turnout rate at every weight level.

The same five DAOs are used for the turnout, concentration, timing, and recruitment comparisons. We define the panel and report these measurements together in “Empirical results from one DAO panel”.

Decentralized vote-market model

Voting and preferences.

There is one binary proposal. Let 𝑥=1 mean passage and 𝑥=0 rejection. Account 𝑖 has fixed snapshot voting weight 𝑤𝑖. An eligible account can vote yes, vote no, or abstain; casting a ballot costs 𝑘𝑖>0. Costs are public in the benchmark. Voting weights, current holdings, and economic exposure are different quantities.

Agent 𝑖 privately observes 𝑢𝑖∼𝐹𝑖, independently across agents, with common knowledge of the distributions and 𝔼|𝑢𝑖|<∞. Here 𝑢𝑖 is its total value of passage relative to rejection. A negative value means passage harms the agent by |𝑢𝑖|; it does not mean the agent has a negative balance. The baseline allows both supporters and dissenters. Agents are risk neutral, and outcome values, costs, and transfers use a common monetary unit.

Decentralized contracts.

Any eligible voter who is not a background voter can make offers to any other eligible voter who is not a background voter. Agent 𝑗 offers 𝑏𝑗𝑖𝑌,𝑏𝑗𝑖𝑁≥0 to voter 𝑖. The offers add across payers. A payment executes only if the recipient casts the specified ballot, regardless of whether that side wins. Abstention receives no payment. Unclaimed escrow is returned after voting closes. Offers are binding and fully funded; for simplicity impose

∑𝑖≠𝑗(𝑏𝑗𝑖𝑌+𝑏𝑗𝑖𝑁)≤𝐵𝑗,

where 𝐵𝑗 is the agent’s available payment budget. Voting fees require separate available funds. This escrow convention is conservative because a recipient cannot cast both ballots.

Let 𝑎𝑖∈{𝑌,𝑁,∅} and set 𝑏𝑗𝑖∅=0. Utility, normalized relative to rejection without participation or transfers, is

𝑈𝑖=𝑢𝑖𝑥−𝑘𝑖𝟏{𝑎𝑖≠∅}+∑𝑗≠𝑖𝑏𝑗𝑖𝑎𝑖−∑𝑗≠𝑖𝑏𝑖𝑗𝑎𝑗.

There is no additive utility baseline. Normalization is harmless for incentives; we do not take a Nash product of these possibly negative utilities.

Timing and solution concept.

In the baseline, agents observe their types, simultaneously commit offers, observe the full offer matrix, and then vote simultaneously. We seek perfect Bayesian equilibria: offers and votes are optimal given beliefs, and beliefs follow Bayes’ rule where applicable. Known 𝐹𝑖 does not determine behavior without costs, contracts, and equilibrium strategies. Below we also study a defender responding to fixed attacker offers, and a defender committing before an attacker responds. These are different timing restrictions, not interchangeable descriptions of the baseline.

Buying abstention

The baseline pays for yes or no. To allow abstention buying, add a nonnegative offer 𝑏𝑗𝑖∅ from payer 𝑗 to voter 𝑖. It is paid after the deadline if the eligible account cast no valid ballot. The utility formula (8) is unchanged, and the escrow constraint now includes abstention offers. The contract needs to verify participation by a snapshot-eligible account, but it need not reveal ballot direction. This lets us model turnout buying and abstention buying under the same ballot secrecy assumption [6], [24].

Let 𝑏𝑌, 𝑏𝑁, and 𝑏0 denote the total offers to an account for yes, no, and abstention. Let 𝑞𝑌, 𝑞𝑁, and 𝑞0 be its conditional passage probabilities under those actions, holding the other voters’ strategies fixed. Abstention is optimal if and only if

𝑏0≥max{𝑏𝑌−𝑘𝑖+𝑢𝑖(𝑞𝑌−𝑞0),𝑏𝑁−𝑘𝑖+𝑢𝑖(𝑞𝑁−𝑞0)}.

Each term on the right is the gain from that ballot over unpaid abstention. With secret direction and verifiable participation, restrict 𝑏𝑌=𝑏𝑁=𝑏. If 𝑝𝑖 is the increase in the chance of the voter’s preferred outcome from casting its preferred ballot rather than abstaining, it strictly prefers to vote exactly when

|𝑢𝑖|𝑝𝑖+𝑏−𝑏0>𝑘𝑖.

Thus paying for abstention raises the turnout threshold by the same amount that an equal payment for either ballot lowers it. Neither payment changes which ballot the voter prefers at fixed beliefs.

Crowd-out. A holder may stop voting even though nobody pays it to abstain. When a subsidy brings in other votes for the same side, the holder’s own vote can become less likely to change the outcome. If its expected benefit then falls below its voting cost, it stays home. This is an equilibrium response to a change in pivotality, separate from the direct incentive created by 𝑏0.

For example, fix two equal-weight unpaid background ballots, one yes and one no, and suppose a weight-2 whale values passage at 3 and a weight-1 holder values it at 1.5 with probability 90% and at zero otherwise. Each ballot costs 1, and a fair coin breaks ties. Without payments, the whale votes yes and the small holder abstains. Offer the small holder 0.5 for either ballot. There is then an equilibrium where the small holder votes yes whenever its value is 1.5 and the whale abstains. With the small holder voting 90% of the time, the whale’s vote raises passage probability by only 0.05, so the whale gains only 0.15 from voting before costs. Among the two voters who can change their ballots, expected turnout falls from one voter to 0.9, and expected cast weight falls from 2 to 0.9. The background ballots are unchanged, so overall turnout and cast weight also fall. This illustrates the response to a given subsidy; it does not claim a buyer would choose it.

Special case: a malicious proposal

An attacker has fixed yes weight 𝐴 and value 𝑉>0 for passage. Every other holder has

𝑢𝑖=−𝐿𝑖≤0,

where 𝐿𝑖 is its private loss from passage. Thus only the attacker intrinsically wants the proposal to pass, although opponents may accept payments for yes. This is a benchmark for an overtly harmful proposal, not an assertion about every participant in a disputed historical vote.

The proposal passes if and only if yes weight strictly exceeds no weight. Ties reject. Snapshot weights remain fixed after bribes are paid. The attacker’s own ballot can suffice for passage if it outweighs the opposition that votes without payments; otherwise the attacker must recruit more yes weight to take the lead.

Why an opponent may need payment

At the voting stage, let 𝑦𝑖 be the total offered payment for yes and 𝑑𝑖 the total for no. Let 𝑞𝑌,𝑞𝑁,𝑞0 be passage probabilities under the voter’s three choices, given its information and other voters’ strategies. Monotonicity of weighted majority gives 𝑞𝑁≤𝑞0≤𝑞𝑌. Its gains from no relative to abstention and yes are

𝑈𝑖(𝑁)−𝑈𝑖(∅)=𝑑𝑖−𝑘𝑖+𝐿𝑖(𝑞0−𝑞𝑁),𝑈𝑖(𝑁)−𝑈𝑖(𝑌)=𝑑𝑖−𝑦𝑖+𝐿𝑖(𝑞𝑌−𝑞𝑁).

These are conditional expected differences; payments to other voters are fixed when comparing its own ballot. Voting costs cancel between yes and no, but not between voting and abstaining.

An opponent can abstain when its effect on passage is too small to justify 𝑘𝑖. Defensive payments cover that gap. Intrinsic opposition also lets it prefer no even when the attacker offers more money. The probabilities in these inequalities depend on the induced voting behavior; they cannot be read directly from 𝐹𝑖.

Paying for turnout or for a no vote?

A turnout payment may be cheaper to administer if participation is easier to verify than ballot direction. The payment needed to recruit an opponent is a different question. Hold other voters’ strategies fixed and suppose there are no competing offers. An opponent with loss 𝐿𝑖 prefers no to abstaining when

𝑏+𝐿𝑖(𝑞0−𝑞𝑁)>𝑘𝑖.

The same condition holds whether 𝑏 pays for either ballot or only for no. In the turnout contract, the opponent still weakly prefers no to yes; suppose it chooses no if indifferent between the two. Thus the minimum offer threshold in both cases is max{0,𝑘𝑖−𝐿𝑖(𝑞0−𝑞𝑁)}. Search, monitoring, or privacy costs can make one contract cheaper in practice, even when the required payment is the same.

If preferences are uncertain, a turnout payment can also mobilize a supporter of the proposal. At a given offer, paying only for no leaves every type that would choose no under the turnout offer willing to do so: it preserves the payoff from no and removes the payment for yes. It may also switch some yes voters to no. This comparison holds for an individual at fixed beliefs about the other ballots; it does not order equilibria or administrative costs.

Which accounts are worth recruiting?

Let 𝐷𝑖 be the expected outcome-value gain from voter 𝑖’s preferred ballot over abstaining, before payments and voting costs. Let 𝜎𝑖 denote that preferred direction, using a fixed rule for ties. Ignoring ties at the turnout cutoff, a turnout offer 𝑏 brings in exactly the types in the event 𝐴𝑖(𝑏)={𝑘𝑖−𝑏<𝐷𝑖<𝑘𝑖}. Its expected increase in the no-minus-yes margin is

𝑤𝑖(Pr(𝐴𝑖(𝑏),𝜎𝑖=𝑁)−Pr(𝐴𝑖(𝑏),𝜎𝑖=𝑌)).

Expected payment is 𝑏Pr(𝐷𝑖+𝑏>𝑘𝑖), including payments to people who would already vote. A large account that reliably votes without an offer may therefore buy less additional opposition per dollar than a smaller, intermittently active delegate. The sign of the expected margin gain depends on the preferences of the newly mobilized types rather than the whole set of recipients.

This is a screening rule. The expected margin alone does not tell us the probability that enough votes arrive together. In “Empirical results from one DAO panel” we compare ranking by snapshot weight with ranking by weight discounted for past turnout, using the same proposals and number of offers.

A blocking coalition

Fix a candidate set 𝑆 of reachable, still-eligible voters. Let 𝑌 bound all yes weight outside 𝑆, and let 𝑁 be fixed no weight outside 𝑆. The set blocks passage if

𝑁+∑𝑖∈𝑆𝑤𝑖≥𝑌.

In this model, the additional no weight required at a given tally is max{0,𝑌−𝑁}. Using a historical final tally for 𝑌 and 𝑁 is a retrospective calculation. For a strategic guarantee, 𝑌 must include or bound future attacker recruitment; it cannot simply be replaced by the votes visible at detection.

Lemma 1 (Procurement against fixed competing offers).

Suppose 𝐿𝑖≥0 for every recipient. Once attacker yes offers 𝑦𝑖 are fixed and observed, protective payments

𝑑𝑖>max{𝑘𝑖,𝑦𝑖},𝑖∈𝑆,

make no strictly preferable to both alternatives for every recipient type. If (16) holds, the coalition blocks passage. The required total defensive payment can be arbitrarily close from above to

𝐶(𝑆,𝑦)=∑𝑖∈𝑆max{𝑘𝑖,𝑦𝑖}.

Proof. Both differences in (12)–(13) are strictly positive under the stated payments, regardless of beliefs about other votes. All recipients vote no, so (16) prevents a strict yes majority. Payment is due even if the proposal would have been rejected anyway. ◻

This guarantee uses only the sign restriction on losses, not distribution shapes. It is stronger, but potentially more expensive, than recruiting voters using beliefs about their pivotality. If actual costs are private, a valid upper bound on each 𝑘𝑖 can replace it; observed gas expenditure alone is not such an upper bound.

Proposition 2 (A profitable protective intervention).

Consider a holder 𝑗∉𝑆 with no other payment obligations or incoming bribes. Hold its own voting action fixed. Suppose a funded intervention satisfying Lemma 1 raises the probability of rejection by Δ𝑃>0, relative to a specified continuation without its offers. If its budget covers total payment 𝑇 and

𝐿𝑗Δ𝑃>𝑇,

then funding the intervention strictly improves its expected utility.

Proof. The intervention increases outcome utility by 𝐿𝑗Δ𝑃 and costs 𝑇. Its own voting cost is unchanged. Their difference is positive. ◻

When the coalition guarantees rejection and the attack otherwise passes certainly, Δ𝑃=1. More generally, Δ𝑃 is a causal comparison between specified strategies with and without the offers, not a correlation between payments and votes. A positive lower bound on 𝐿𝑗Δ𝑃 suffices for the comparison.

No designated defender is assumed: any holder satisfying these conditions can fund the coalition. However, Proposition 2 does not establish an equilibrium in which it does. Others may fund first, free-ride, or make competing offers. In the simultaneous-offer baseline, it rules out a proposed equilibrium only when the intervention is a feasible profitable unilateral deviation, with the other offer strategies held fixed. A response to newly observed attacker bids is not a simultaneous-stage deviation.

Simultaneous competing offers

We first isolate the funding problem. There is one pivotal voter: the proposal passes if this voter abstains or votes yes, and fails if it votes no. The other ballots, including a positive-weight unpaid background ballot, are fixed. The pivotal voter is neutral about passage and pays 𝑘>0 to vote. The attacker values passage at 𝑉>0 and offers 𝑎∈[0,𝐵𝐴] for yes. Each defender 𝑗 loses 𝐿𝑗≥0 from passage and offers 𝑑𝑗∈[0,𝐵𝑗] for no. All funders are eligible voters. Offers are simultaneous, additive, and continuous. Only offers for the ballot cast are paid, and unused escrow has no carrying cost. The pivotal voter resolves payoff ties in the order no, abstain, yes.

Theorem 3 (Simultaneous funding). A protective pure Nash equilibrium exists if and only if

∑𝑗min{𝐵𝑗,𝐿𝑗}≥max{𝑘,min{𝐵𝐴,𝑉}}.

An equilibrium in which the proposal passes exists if and only if no defender has both 𝐵𝑗≥𝑘 and 𝐿𝑗>𝑘.

Proof. Let 𝐷=∑𝑗𝑑𝑗. The pivotal voter earns 𝐷−𝑘 from no, 𝑎−𝑘 from yes, and zero from abstention. Therefore no is chosen exactly when 𝐷≥max{𝑘,𝑎}. At a protective equilibrium no defender pays more than its loss: withdrawing its offer saves the payment and costs at most 𝐿𝑗. If 𝐷<min{𝐵𝐴,𝑉}, the attacker can outbid 𝐷 at a price below 𝑉, gaining strictly. Thus 𝐷 must cover 𝑘 and min{𝐵𝐴,𝑉}, and the individual loss and budget constraints give the bound.

Conversely, let the attacker offer 𝑎=min{𝐵𝐴,𝑉} and split 𝐷=max{𝑘,𝑎} among defenders within their individual limits. The voter chooses no. Cutting any positive contribution causes passage and saves at most its contributor’s loss. Raising a contribution is costly and leaves the outcome unchanged. The attacker cannot profitably offer more than 𝐷. This proves existence.

If the proposal passes at an equilibrium, the pivotal voter cannot vote yes: that requires 𝑎>max{𝑘,𝐷}, so the attacker could shave its offer while preserving passage. The voter must abstain, so 𝐷<𝑘 and 𝑎≤𝑘. Any defender with 𝐵𝑗≥𝑘 and 𝐿𝑗>𝑘 could then secure rejection on its own by offering 𝑘. Conversely, if no such defender exists, zero offers are an equilibrium because passage is free for the attacker and no defender can profitably cover 𝑘 alone. □

For example, take 𝑘=5, 𝑉=𝐵𝐴=6, and two defenders with 𝐵𝑗=𝐿𝑗=4. A defense funded by two offers of 3 and zero offers are both equilibrium outcomes (the attacker offers 6 in the protective equilibrium). Pooling enough money is compatible with a failure to coordinate.

This is an existence result in ordinary pure Nash equilibrium. If 𝐵𝐴≥𝑉>𝑘, every protective pure equilibrium uses a weakly dominated attacker offer: 𝐷≥𝑉, and minimizing the defenders’ payments requires 𝐷=max{𝑘,𝑎}, so 𝑎≥𝑉. An offer that high never yields a positive payoff. An offer of zero yields a nonnegative payoff, and yields 𝑉 whenever 𝐷<𝑘 because the proposal passes if the pivotal voter abstains. This refinement depends on the default outcome and the preference over tied offers; the theorem does not give existence for arbitrary weighted electorates.

If the attacker responds

Now the defense commits first. It gives each recruit 𝑖∈𝑆 a binding no offer 𝑑𝑖>𝑘𝑖. A disjoint set 𝐼 of reachable voters has no offers. Other ballots are fixed, with yes weight 𝑌0 and no weight 𝑁0, including the unpaid background ballot. For a set 𝑇, write 𝑊(𝑇)=∑𝑖∈𝑇𝑤𝑖. The provisional no lead if 𝑆 votes no and 𝐼 abstains is Δ=𝑁0+𝑊(𝑆)−𝑌0. Assume Δ≥0. The attacker can offer payments for yes to voters in 𝑆 or 𝐼.

Theorem 4 (Cost of breaking a committed defense). For neutral recipients who choose yes at a payoff tie, the minimum funded counteroffer that restores passage is the minimum of

∑𝑖∈𝐹𝑑𝑖+∑𝑗∈𝑈𝑘𝑗

over 𝐹⊆𝑆 and 𝑈⊆𝐼 such that

2𝑊(𝐹)+𝑊(𝑈)>Δ.

If no such pair exists, the cost is infinite. Finite budget 𝐵𝐴 can break the defense exactly when it covers this minimum. An attacker with value 𝑉 strictly gains from a successful response if and only if it can afford the minimum and 𝑉 exceeds it.

Proof. Flipping a recruit in 𝑆 to yes costs at least its no offer 𝑑𝑖. Activating an otherwise inactive voter in 𝐼 costs at least its voting cost 𝑘𝑗. A flip adds 𝑤𝑖 to yes and removes 𝑤𝑖 from no; activation adds 𝑤𝑗 to yes. Thus every successful response has a pair of sets satisfying the constraint and pays at least its cost. Conversely, offer exactly 𝑑𝑖 to each selected recruit and 𝑘𝑗 to each selected inactive voter. The specified tie rule yields precisely those yes ballots, so the resulting strict yes majority restores passage at the claimed cost. □

The same minimum is a worst-case guarantee against privately weak opponents of passage, provided only the specified recipients can change their ballots. Each recruit still strictly prefers no to abstaining. Neither a recruit offered less than 𝑑𝑖 nor an uncontracted voter offered less than 𝑘𝑖 can prefer yes. Extra opposition from the latter set makes passage harder, and neutrality attains the bound. With tie rules that favor no or abstention, the affected prices are strict thresholds, so the expression gives an infimum and equality of the budget may not be sufficient.

A larger total defensive budget alone is insufficient. Suppose the attacker has 1.5 yes weight, an unpaid no ballot has weight 0.25, and there are two unit-weight recruits. Each has voting cost 1 and a defensive offer of 3.5. Flipping either recruit reverses the tally and can cost less than the defense’s total of 7. Conversely, when the defense has enough spare weight to survive a flip, the attacker may have to outbid several offers. Repeated counterbidding is outside this model.

Empirical conditions for a defense

The target is an audited decision for which capacity, timing, funding incentives, and counteroffer robustness hold together. Showing that large holders exist or votes can be bought establishes only part of the argument.

ConditionEvidence needed
Blocking capacityHistorical rules and snapshot weights; exclude unusable/already-cast votes; bound future yes recruitment.
Timely executionDocumented detection time, remaining voting window, contact/signing delays, and usable payment contracts.
Affordable recruitmentCredible upper bounds on total participation costs or observed acceptance prices; include operational costs beyond gas.
Private incentiveVerified economic exposure, a justified avoided-loss estimate, and a lower bound on improvement in rejection probability.
Available fundingA holder’s liquid budget or already committed decentralized funding; DAO treasury value is not a private budget.
Counteroffer robustnessObserved binding attacker offers or a defensible spending bound under an explicit timing protocol.

Where inputs are uncertain, use conservative bounds: 𝐿𝑗≥𝐿𝑗, Δ𝑃≥𝛿, and recruitment cost at most 𝐶. The funding condition is certified by

𝐵𝑗>𝐶,𝐿𝑗𝛿>𝐶,

provided capacity and timely execution are separately established. Bounds must come from evidence or be labeled sensitivity assumptions. In particular, a token-price change around an attack does not by itself identify the damage caused by it. Learning 𝐹𝑖 or a causal response to payments needs incentive episodes, an experiment, or additional identifying assumptions.

The common panel measures capacity and voting behavior. It contains no verified schedules of competing offers, reservation prices, private losses, or voter-level defensive transfers. We do not estimate outcomes that require those inputs.

Empirical results from one DAO panel

Sample. The panel is Compound, Uniswap, ENS, Gitcoin, and Nouns, from the covered Ethereum governor deployments through 31 October 2026 UTC (block 26,316,188). These are the five DAOs in our prior data collection for which we have complete delegation checkpoint and ballot histories. Selection rests on data coverage. The DAOs are established Ethereum projects with different governance uses and electorate sizes, but are not a probability sample of DAOs or blockchains. Every panel analysis uses these same five DAOs. The unit is a proposal on a named governor; a resubmission has a new proposal ID.

Of 1,698 submissions, we exclude 189 canceled or vetoed by the scheduled deadline, one whose window had not completed, and eight with no positive-weight yes or no ballot. We retain 1,500 and reconcile every positive ballot weight against its eligible snapshot weight. We include failed proposals in the turnout, concentration, and timing analysis. To ask how often additional opposition could change a passing tally, we restrict that analysis to the 1,062 proposals that were in a passing state at the cutoff or had an execution event. The ballot screen cannot verify the theoretical assumption that a ballot was unpaid.

Turnout, concentration, and timing

DAOProposalsAddresses votingPower castTop 1% voting
Compound3613.0%28.6%22.0%
Uniswap864.0%27.4%9.9%
ENS761.1%32.3%25.2%
Gitcoin1150.8%33.0%7.0%
Nouns86211.8%19.1%34.4%

Table 1: Address and power turnout are proposal averages. Positive-weight explicit abstentions count as turnout. The last column pools address–proposal pairs within the top one percent of eligible snapshot weights, rounding the number of addresses up.

Power turnout exceeded address turnout in 1,146 of 1,500 proposals. The mean gap was positive in every DAO. The five largest voting addresses accounted for more than half the cast weight in 1,337 proposals; one address alone did so in 94. Positive-weight yes/no address counts and yes/no weight gave opposite strict majorities in 99 proposals. In 59, the yes-minus-no margin had opposite signs at 90% of the scheduled voting blocks and at close. Nouns accounts for 91 of the majority disagreements and 56 of the late reversals. This concentration of events is a reason to compare DAOs separately.

In the 1,009 proposals with a positive-weight no ballot, remove the largest no ballot and check whether reinserting it is decisive. Under the majority rule with ties rejecting, adding opposition equal to one percent of eligible weight made that incumbent pivotal in 31 cases and removed its pivotality in 21 cases. This holds the other ballots fixed and conditions on an observed opponent. It says neither that the largest opponent would have changed its turnout nor that the added votes would have been available at a given price.

Can ten targets erase the yes lead?

At each passing proposal’s scheduled block midpoint, rank positive-weight snapshot addresses that have not yet voted. Keep observed ballots fixed. Select up to ten addresses, and give every selected final nonvoter a hypothetical no ballot. The primary rule ranks by snapshot weight alone. A second rule multiplies weight by estimated nonparticipation: use the address’s record among proposals whose windows closed before the current snapshot, with one prior vote and one prior nonvote. Both rules count a tie with the recorded yes weight as enough to erase the yes majority; this is a sufficient route to rejection and need not be the cheapest one.

DAOPassingBy weightShareBy discounted weight
Compound31726082.0%270
Uniswap786684.6%69
ENS702434.3%23
Gitcoin976263.9%63
Nouns50035170.2%347

Table 2: Counts of passing or executed proposals whose yes lead the selected ten could erase under full compliance. An address that eventually voted contributes zero additional weight.

The weight rule covers 763 of 1,062 proposals (71.8%); giving each DAO equal weight gives 67.0%. Leaving out any one DAO gives 67.5% to 74.5%. The discounted rule covers 772 of the same 1,062 (72.7%; 68.2% with equal DAO weights). Its gain is small and varies by DAO.

Random targeting benchmark. Keep the elections and turnout records fixed and draw up to ten addresses uniformly without replacement from each proposal’s actual midpoint pool. Simulate 1,999 independently drawn panels using this rule. The mean fraction covered, weighting each DAO equally, is 1.3%, versus 67.0% for the weight rule. No draw reached that observed rate, giving a Monte Carlo tail probability of 0.0005 with the add-one correction under the uniform-targeting benchmark. The randomization is over the hypothetical targets. It supplies no confidence interval or significance test for a population of DAOs.

Sensitivity to compliance. If each selected final nonvoter independently accepts a no offer with probability 75%, the weight rule blocks 54.8% in expectation with equal proposal weights and 46.9% with equal DAO weights. If acceptance is correlated but each individual probability is at least 75%, the corresponding conservative lower bounds are 32.6% and 25.1%. In a different comparison, suppose all selected final nonvoters vote but each supports no with probability 75%. The net no weight is then half as large in expectation: the expected margin crosses the blocking threshold in 390 cases for the weight rule and 395 for the discounted rule. A margin in expectation is not a probability of blocking. None of these scenarios measures real acceptance.

Scope and inference. The natural unit for generalizing across governance systems is the DAO or chain, not the individual proposal. Repeated proposals share delegates, institutions, and market conditions. With five purposively selected DAOs on one chain, proposal-level standard errors would overstate the evidence about other DAOs. To estimate a cross-chain rate, first fix a dated registry of active DAO governors and define the target as active DAOs with verifiable token-weighted binary decisions. Sample across chain families, and within them across DAO size and governance type. Record reconstruction failures before looking at results, report coverage by stratum, and use the same included DAOs for each outcome. Predeclare the primary measure and the minimum effect worth detecting, then size the sample for uncertainty across DAOs and chains. Even that design only generalizes to the defined class of governance systems.

Historical illustration: Compound Proposal 289

This case was selected to show what an investigation can recover. Proposal 289 was canceled after the vote, so it is not in the 1,062-proposal passing subset. It is included in the 1,500 completed votes; the extra case-specific calculations below are separate.

A Compound forum warning identified Proposal 289 at 16:02 UTC on 25 July 2024. This is a documented public alert, not necessarily the first privately known warning. The voting snapshot was at 23:43 UTC that day (block 20,386,918), and the last voting block was at 17:45 UTC on 28 July (block 20,406,628). That left about 66 hours from the snapshot. One hour before close, yes trailed 118,600 to 583,637; about 82.6% of the final yes weight arrived in the last hour. The proposal finished 48,554.70 COMP ahead, met the governor’s requirements for passage, and was canceled before execution [16], [18].

Capacity available for recruitment. Rank all snapshot-eligible addresses by voting weight before any ballot on this proposal has been cast. The two largest have 256,007 and 101,008 COMP of voting power. Neither subsequently voted. Against the recorded final tally, either would have been sufficient by itself; the largest alone gives a hypothetical no lead of 207,452 COMP, and the pair gives a lead of 308,460 COMP. This uses an address ranking that could have been formed at the snapshot. It uses the final tally only to evaluate that ranking. The records do not show that either address was contactable or willing to vote no.

Competition for the same voters. Apply Theorem 4 while holding the historical ballots fixed. Give each selected recruit a binding no offer of one monetary unit, assume recipients are neutral, and set a common cost for new yes votes from the remaining final nonvoters. These are price scenarios. The results assume every such address can be reached; they are not a forecast or an estimate of the attacker’s actual budget.

Price to activate a new yes voterMinimum counteroffer with one recruitMinimum counteroffer with two recruits
0.1 unit0.3: three new yes votes0.6: six new yes votes
0.5 unit1: flip one recruit1: flip one recruit
1 unit1: flip one recruit1: flip one recruit

Table 3: Counteroffers in monetary units, at one unit per defensive recruit. With two recruits the largest can still be decisive: flipping it swings the margin by about 512,014 COMP.

Costs and funding. The 58 recorded ballot events came from 55 distinct transactions. Their median transaction fee was 0.000188 ETH (range 0.000094 to 0.001181 ETH). Some transactions carried more than one ballot. The fee is a transaction cost, not a voter’s reservation price or the cost of arranging a payment. Consider an assumed offer of 0.1 ETH to each recruit. The proposer of the defense must be able to fund 0.1 ETH for one recruit or 0.2 ETH for two, excluding any separate operating costs. At an assumed increase in rejection probability of one half, the break-even private losses are 0.2 ETH and 0.4 ETH, respectively. Of the 20 addresses that eventually cast a positive-weight no ballot, 10 held at least 0.1 ETH and eight held at least 0.2 ETH at the snapshot. This is an address-level balance check on observed opponents, not proof that their controllers could or would spend those funds. We do not observe binding offers or any stakeholder’s actual available budget or private loss. The requested transfer of 499,000 COMP is the size of the proposed allocation, not the loss to a particular voter or the harm actually realized. The proposal’s eventual cancellation also means that preventing passage cannot be equated with preventing an executed transfer.

When do payments equalize holdings?

Fix a set of accounts and one asset. Isolate the transfers made by the defense, holding fees, treasury changes, and all other transfers fixed. Keep accounts in the comparison even if their balance falls to zero. Holdings in the asset paid and voting weights at the proposal snapshot are different objects.

Proposition (One transfer). If account 𝑗 starts with balance 𝑠𝑗>𝑠𝑖, a positive transfer 𝑡≤𝑠𝑗 to account 𝑖 makes the distribution strictly more equal in the majorization order if and only if

0<𝑡<𝑠𝑗−𝑠𝑖.

At 𝑡=𝑠𝑗−𝑠𝑖 it just exchanges the balances; a larger transfer makes the distribution strictly less equal.

Proof. The pair’s total is constant and its absolute gap after payment is |𝑠𝑗−𝑠𝑖−2𝑡|. The gap shrinks exactly over the stated interval; equivalently, the new pair is a nontrivial convex combination of the original pair and its swap. The same equalizing comparison holds with the other balances fixed. □

For many transfers, the exact test is to sort the balance vectors before and after, then compare the sums of their largest 𝑚 entries for every 𝑚. Holdings become more equal if and only if none of those sums increases and at least one falls, with total holdings fixed. A simple sufficient condition for weak equalization is that a single sponsor’s balance after all payments is at least every recipient’s starting balance. Apply the one-transfer test in sequence using the sponsor’s running balance: no payment exceeds the gap at the time it is made.

This gives a reason to expect equalization when a large owner pays smaller owners. It does not follow from successful defense: a delegate can have large voting weight with little personal exposure, and a payment in ETH does not rebalance COMP holdings.

What the transfer records can establish

To apply the test, identify the actual payer, recipients, payment asset, amounts, and balances just before payment, using the same account set on both sides. The panel has no verified voter-level defensive payment records, so we report no estimate of how defense affects holdings. An isolated purchase outside the panel cannot fill that gap.

What remains to prove and validate

The current positive results are conditional procurement and incentive statements. A profitable intervention alone is not an equilibrium result.

Pending: determine when protected equilibria exist in general weighted vote markets, and which guarantees survive equilibrium multiplicity and attacker responses.

The historical replays identify where a protective coalition had enough unused snapshot weight and where opposition already defeated an attack. No reconstructed case establishes protective bribery in equilibrium: that requires evidence about who would accept an offer, who would fund it, and whether an attacker could respond. The holdings claim separately requires transfers tied to the voters who ultimately received them.

Costly turnout [25], competing vote buyers [26], and voluntary public-goods funding [27] provide the relevant starting points. The intended contribution is to connect these incentive obstacles to fixed-eligibility governance attacks and testable conditions for a decentralized defense.

References