Worked example · 2024-07

Compound DAO Proposal 289 attempted governance takeover — Ethereum — 2024-07-28

Loss: $0 — but ~$24M of COMP would have been transferred from the Compound DAO's treasury to a "Golden Boys"-controlled multisig / yield-bearing vault if Proposal 289 had been allowed to execute. The proposal passed the on-chain vote on 2024-07-28 with the Golden Boys cohort accumulating sufficient COMP voting power to clear quorum; the transfer was prevented in the pre-execution review window when the broader DAO recognised the proposal as a self-dealing transfer of treasury COMP into a vehicle effectively controlled by the proposing cohort. An emergency counter-proposal was submitted and the original proposal was effectively cancelled / superseded before the timelock-gated execution window opened the treasury action for execution.
Recovery: full — no funds were actually transferred. The case is recorded in OAK as an attempted but prevented T9.003 governance attack, structurally important for the corpus precisely because the timelock-gated governance lifecycle worked as designed: the on-chain vote passed, but the pre-execution review window was sufficient to recognise the self-dealing intent and deploy a counter-proposal before the timelock expired and the malicious payload became executable.
OAK Techniques observed: OAK-T9.003 (Governance Attack) — primary, in the accumulated-voting-power sub-class: the attacking cohort acquired sufficient COMP voting power through direct on-chain accumulation across multiple addresses over a multi-week window prior to proposal submission, without flash-loan funding. This is the load-bearing structural distinction from Beanstalk 2022 (/examples/2022-04-beanstalk.md), which is the canonical T9.003 case in the flash-loan-funded sub-class. The two together span the canonical sub-classes for T9.003 voting-power acquisition. OAK-T16.003 (Delegation-Cluster Vote Takeover — operational sub-pattern of T9.003 where the load-bearing acquisition mechanism was multi-wallet delegation pull from the publicly-identified Golden Boys / Humpy cohort); OAK-T16.002 (Hostile-Vote Treasury Drain — co-occurring sub-pattern, since the cohort's direct COMP holdings also contributed to the vote weight alongside the delegation pull); the case is the canonical T16.003 + T16.002 *prevented*-attempt anchor.
Attribution: pseudonymous — publicly named cohort, individuals pseudonymous-to-identified. The acquiring cohort was publicly identified at the time as the "Golden Boys" — a coordinated group of COMP accumulators including the wallet associated with the on-chain identity "Humpy" (also publicly active in similar patterns at Balancer, SushiSwap, and other governance forums). The cohort identification is on the public DAO-forum record and was the basis on which the broader Compound community recognised the proposal as self-dealing within the pre-execution review window. There is no civil or criminal action — the attempt was prevented before any treasury extraction.
Key teaching point: Compound 2024 is OAK's canonical worked example for the accumulated-voting-power sub-class of T9.003, paired with Beanstalk 2022 (flash-loan-funded sub-class) as the cohort canonical for T9.003 voting-power acquisition sub-classes. The two together span the canonical sub-class structure: realised vs prevented outcome, single-transaction vs multi-week accumulation timeline, T9.002 chain vs no-precondition framing, total-loss vs full-recovery state.

Summary

Compound is the canonical EVM lending protocol on Ethereum, with on-chain governance via the Governor Bravo contract pattern. COMP is the protocol's governance token; voting power is computed from snapshotted (checkpointed) COMP balances at proposal-submission time, with a timelock between vote-passage and execution that opens a defender review window before privileged actions become executable.

In the months preceding July 2024, a coordinated cohort publicly identified as the "Golden Boys" — including the wallet associated with the on-chain pseudonymous identity "Humpy" — accumulated COMP across multiple addresses through direct on-chain purchases. By late July 2024 the cohort's accumulated COMP voting power was sufficient to clear the Compound governance quorum threshold, particularly when amplified by delegations from supporting holders.

On 2024-07-28 the cohort submitted Compound Proposal 289, which directed the Compound DAO treasury to transfer approximately 499,000 COMP (~$24M at the price of the day) into a "goldCOMP" yield-bearing vault that was effectively controlled by the proposing cohort. The proposal passed the on-chain vote — the cohort's accumulated voting power, plus delegations from supporting holders, cleared quorum and the simple-majority threshold.

Between vote-passage and the timelock-gated execution window, the broader Compound community — through DAO-forum discussion and on-chain analytics partners — recognised the proposal as a self-dealing transfer of treasury COMP into a vehicle controlled by the proposing cohort, rather than a routine treasury-management action of the kind the proposal text framed it as. An emergency counter-proposal (and supporting community actions, including pressure on the proposing cohort itself) was deployed to cancel / supersede Proposal 289 before the timelock expired and the malicious treasury action became executable. The cohort ultimately participated in the cancellation and a renegotiated, scaled-back arrangement was eventually settled at substantially reduced scale and under different governance terms.

For OAK's purposes the case is the canonical 2024 worked example for the accumulated-voting-power sub-class of T9.003, structurally distinct from the flash-loan-funded sub-class canonical at Beanstalk 2022. The case is also notable as the canonical OAK example of a prevented rather than realised T9.003 attack — the timelock-gated governance lifecycle worked as designed, and the case documents the operational shape of a successful pre-execution defence.

Timeline (UTC)

When	Event	OAK ref
Pre-2024	Compound governance operated through Governor Bravo with snapshotted COMP voting power and timelock-gated execution; standing T9.003 surface gated by the timelock review window	(T9.003 surface — design-level, latent)
2024 (months prior)	"Golden Boys" cohort begins coordinated on-chain accumulation of COMP across multiple addresses; pattern publicly visible on Compound governance forum and on-chain analytics dashboards (Tally, Boardroom, etc.)	(T9.003 setup — voting-power accumulation)
2024-Q2 / 2024-Q3	Cohort accumulated COMP voting power approaches and clears quorum-relevant threshold; community awareness of cohort coordination grows	(T9.003 setup, cohort visibility)
2024-07-28	Compound Proposal 289 submitted by Golden Boys-aligned proposer; proposal text frames the action as routine treasury yield-deployment to "goldCOMP" vault; on-chain payload directs ~~499,000 COMP (~~$24M) from Compound DAO treasury to the cohort-controlled vault	T9.003 execution attempt — proposal submission
2024-07-28 (vote period)	Voting period elapses with the cohort's accumulated voting power plus delegations clearing quorum and simple-majority threshold; Proposal 289 passes the on-chain vote	T9.003 — vote phase complete
2024-07-28 — pre-execution timelock	Compound community recognises the proposal as self-dealing self-allocation of treasury COMP to cohort-controlled vehicle; DAO-forum discussion escalates rapidly	(defender response — pre-execution review)
Pre-execution timelock expiry	Emergency counter-proposal / cancellation action deployed before timelock-gated treasury action becomes executable; proposal effectively cancelled / superseded	T9.003 execution prevented
2024-08 onward	Renegotiated, scaled-back arrangement settled at substantially reduced scale and under different governance terms; cohort participates in the cancellation	(recovery state — full)
Continuing	Case documented in industry retrospectives as the canonical 2024 attempted but prevented governance attack on a top-tier DeFi protocol	(cohort signal)

What defenders observed

The timelock-gated governance lifecycle is the load-bearing defensive primitive, and Compound 2024 is the canonical 2024 evidence that it works when used as designed. Beanstalk 2022 is the canonical T9.003 case where the protocol's emergencyCommit function bypassed the normal timelock and made the malicious proposal immediately executable upon vote-passage; ~$182M was extracted from the treasury before the community could respond. Compound's Governor Bravo design retains the timelock between vote-passage and execution as a mandatory review window. In the 2024 case, the on-chain vote passed but the timelock expired without the malicious action being executed because the community deployed a counter-proposal in the review window. The defender lesson is that the timelock is not redundant or annoying — it is the load-bearing mitigation that converts a passed-vote governance attack from a realised loss into a prevented attempt.
Accumulated-voting-power is a structurally distinct T9.003 sub-class from flash-loan-funded voting-power. Beanstalk 2022 acquired ~67% of STALK voting power within a single transaction window via flash loans; the cohort exited the position the moment the malicious proposal executed. The Golden Boys cohort acquired COMP voting power across weeks-to-months of on-chain accumulation, retaining the position throughout the vote and into the post-cancellation renegotiation. The two sub-classes have materially different defender-detection surfaces:
- The flash-loan-funded sub-class is detectable in real-time by monitoring governance-token-transfer events for sudden large acquisitions correlated with active proposals.
- The accumulated-voting-power sub-class is detectable only through cohort-level surveillance of governance-token accumulation patterns over multi-week windows; the Compound 2024 case was visible in this surveillance precisely because the Golden Boys cohort was publicly known and tracked by DAO-analytics dashboards (Tally, Boardroom, etc.) for some time before Proposal 289 was submitted.
Cohort visibility is a defender-side advantage but not a substitute for governance-design discipline. The Golden Boys cohort's coordination was publicly visible for months; Humpy's pattern of accumulating governance tokens for self-dealing or rent-extracting proposals was previously documented at Balancer and SushiSwap. The Compound community's ability to recognise Proposal 289 as self-dealing within the pre-execution review window depended on this prior visibility. The defender lesson generalises: cohort-level surveillance of governance-token accumulation patterns, particularly by publicly-identified actors with prior pattern-of-conduct, is a discoverable defender practice that pairs with the timelock-gated lifecycle to make T9.003 attempts recognisable and counterable.
Emergency counter-proposal / cancellation tooling is a discrete defensive primitive that the Compound case validates. The successful prevention of Proposal 289 depended on the operational availability of a fast-path counter-proposal mechanism that could be deployed within the pre-execution timelock window. The defender lesson is that the timelock window is not self-defending — it requires an operational counter-proposal capability to be useful. Protocols that have a long timelock but no operational community-level capacity to deploy a counter-proposal in the review window are fragile in a way Compound, with its active community and cohort-aware DAO process, was not.
The case is the canonical 2024 evidence that top-tier DeFi governance has materially hardened since Beanstalk. Post-Beanstalk governance contracts at top-tier protocols adopted snapshot-based voting power and mandatory timelocks as standard. The Compound 2024 case shows that the design-level mitigations work: a passed-vote governance attack on a top-tier protocol was prevented in the timelock window through community action. Smaller-protocol T9.003 incidents remain common and are typically not published forensically; the cohort framing is that top-tier DeFi has structurally moved past the realised-T9.003 era while smaller protocols continue to fire.

What this example tells contributors writing future Technique pages

T9.003 has two structural sub-classes — flash-loan-funded and accumulated-voting-power — and the 2022/2024 worked-example pair canonicalises them. Beanstalk 2022 is the canonical flash-loan-funded sub-class case (T9.002 + T9.003 chain, single-transaction execution, ~$182M realised loss, no recovery). Compound 2024 is the canonical accumulated-voting-power sub-class case (no T9.002 precondition, multi-week accumulation, ~$24M attempted but prevented, full recovery). The T9.003 technique page should reflect this two-sub-class structure explicitly and avoid implying that flash-loan-resistance alone closes the surface.
Realised vs prevented is a first-class outcome distinction for T9.003 that the corpus should make explicit. Beanstalk 2022 is realised; Compound 2024 is prevented. The two together establish the operational difference between governance designs without timelock-gated execution and governance designs with timelock-gated execution operated by a community with cohort-aware DAO process and counter-proposal capability. Worked examples for T9.003-class cases should explicitly tag the outcome (realised / prevented) in the loss header to make the corpus-level picture legible.
Cohort-level surveillance of governance-token accumulation is a discoverable defender practice that pairs with the timelock-gated lifecycle. Defender-side monitoring tooling (Tally, Boardroom, on-chain analytics dashboards) makes governance-token accumulation patterns publicly observable. Cohort-level surveillance of publicly-identified actors with prior pattern-of-conduct (Humpy at Balancer, SushiSwap, Compound) is a discoverable practice that the Compound 2024 case validates. Contributors writing future T9.003 worked examples should treat cohort-level surveillance as a first-class Mitigation reference.
Emergency counter-proposal capability is a separate defensive primitive from timelock-gated execution and should be documented as such. Timelock-gated execution opens the review window; counter-proposal capability is what fills it. Worked examples should make the distinction explicit: a long timelock with no operational counter-proposal capability is structurally fragile in a way a long timelock with active counter-proposal capability is not.
Publicly-named cohort attribution is a distinct attribution status from pseudonymous and from confirmed-state-actor. The Golden Boys cohort is publicly named; individual members are pseudonymous-to-identified. This is a meaningful attribution status that the OAK actor framework should accommodate alongside pseudonymous and the OAK-Gnn confirmed-state-actor patterns.

Public references

[compoundproposal289_2024] — On-chain record of Compound Proposal 289, including the proposal payload, voting outcome, and subsequent cancellation / supersession actions.
[compoundforumproposal289_2024] — Compound governance-forum thread covering the proposal submission, community recognition of the self-dealing pattern, and the deployment of the counter-proposal in the pre-execution review window.
[goldenboyscompound2024] — Industry retrospective covering the Golden Boys cohort's accumulation pattern, the cohort's prior pattern-of-conduct at Balancer and SushiSwap, and the Compound 2024 attempt.
[blocksecgovernance2024] — BlockSec / industry-side governance-attack retrospective covering the 2024 cohort context and the Compound case as the canonical accumulated-voting-power sub-class example.
[tallygovernancecompound2024] — Tally / Boardroom on-chain governance analytics covering the Compound 2024 voting period, the cohort coordination pattern, and the post-cancellation outcome.

Citations

[compoundproposal289_2024] — On-chain proposal record; primary source for proposal payload, voting outcome, and timelock / cancellation timeline.
[compoundforumproposal289_2024] — DAO-forum record; primary source for community recognition timeline and counter-proposal deployment.
[goldenboyscompound2024] — Cohort attribution; primary source for the Golden Boys cohort identification and prior pattern-of-conduct cross-reference.
[blocksecgovernance2024] — Governance-attack retrospective; primary source for the accumulated-voting-power sub-class framing.
[tallygovernancecompound2024] — On-chain governance analytics; secondary source for cohort coordination pattern.
[zhou2023sok] — academic taxonomy classifying governance attacks as a recurring class.

Discussion

Compound 2024 is OAK's canonical worked example for the accumulated-voting-power sub-class of T9.003, paired with Beanstalk 2022 (flash-loan-funded sub-class) as the cohort canonical for T9.003 voting-power acquisition sub-classes. The two together span the canonical sub-class structure: realised vs prevented outcome, single-transaction vs multi-week accumulation timeline, T9.002 chain vs no-precondition framing, total-loss vs full-recovery state.

The case is the canonical 2024 evidence that the design-level mitigations developed in the wake of Beanstalk — snapshot-based voting power, mandatory timelock-gated execution, cohort-aware DAO process — work as designed at top-tier DeFi protocols. The on-chain vote passed; the malicious treasury action was prevented in the pre-execution review window through community recognition and counter-proposal deployment. The defender lesson is that the timelock is the load-bearing mitigation, and the timelock window is operationally useful only when paired with active counter-proposal capability and cohort-level surveillance of governance-token accumulation patterns.

The publicly-named-cohort attribution status is structurally important for the OAK actor framework. The Golden Boys cohort is not pseudonymous in the standard sense — the cohort's coordination is publicly visible on the DAO-forum and on-chain analytics record, and the cohort includes wallets with prior pattern-of-conduct at other governance forums (Humpy at Balancer, SushiSwap). The cohort's individual members are pseudonymous-to-identified at varying levels. This is a meaningful attribution status that sits between pseudonymous and confirmed in the OAK actor framework and that the framework should accommodate explicitly.

For OAK's broader credibility, including Compound 2024 in v0.1 closes two gaps: it adds the canonical accumulated-voting-power sub-class case to a T9.003 corpus that otherwise has only the Beanstalk flash-loan-funded sub-class case, and it adds the canonical prevented-attempt case to a corpus that otherwise records only realised losses. The structural lesson — that timelock-gated execution paired with active community counter-proposal capability and cohort-level surveillance is a defensible governance design — is downstream of the case and is the cross-cutting OAK contribution. Contributors writing future T9.003 worked examples should reference Compound 2024 as the canonical accumulated-voting-power sub-class case and the canonical prevented-attempt case for the corpus.