Is a DeFi Protocol Decentralized or Just a Multisig

Most protocols described as decentralized can still be upgraded, paused or frozen by a small set of keys. The useful question for a depositor is therefore not whether such keys exist, since they nearly always do, but how many signatures they need, who holds them, and how much warning users get before they are used. Arbitrum, one of the better-documented cases, puts the number at nine. Nine of its twelve Security Council members can act in an emergency, and in April 2026 they did, moving 30,766 ETH (about $71 million) out of an exploiter's address.

This page is a checklist that can be run on any protocol from its documentation and governance forum. A protocol that answers all six questions clearly is not necessarily safe. A protocol that cannot answer them is not decentralized, whatever its homepage says.

The mechanics of proposals and votes are covered separately in the token vote and forum process. This checklist concerns the powers that sit outside that process.

Arbitrum's own governance documentation answers most of the six questions directly.

Owner. Chain ownership formally rests with ArbitrumDAO, which can also vote to change the governance system itself.

Threshold. The Security Council has 12 members, and 9 signatures are needed for emergency action.

Signers. ARB holders and delegates elect the members each year.

Delay. For certain governance changes, users get several weeks to withdraw funds to Ethereum before the change takes effect.

Powers without a vote. This is where the documentation is most candid. The same 9-of-12 body can push routine upgrades, such as minor bug fixes, without a full DAO vote. The documentation also concedes that if 9 members were compromised, "the system and users' funds could be compromised."

Override. The DAO elects the council and can rewrite the system, but only on the DAO's own schedule. That schedule is slower than nine people with signing devices.

The verdict is not that Arbitrum is decentralized or that it is not. It is that the answer is written down, and it comes to nine people, all elected. The documentation is marked "public preview," so the details may change, which is a reason to check it again rather than rely on this summary.

Aave. Aave runs two separate guardians. Its Governance Emergency Guardian is a 5-of-9 multisig that can cancel governance payloads that are "malicious, materially erroneous, or otherwise unsafe." A signer rotation proposed on August 12, 2026 keeps that threshold but will disclose only the signing addresses, not who holds them. The stated reason is to reduce phishing and coercion. The cost is that the independence question from the checklist can no longer be checked from outside. A separate Protocol Emergency Guardian is being set to 4-of-7. On top of these, a September 2026 Snapshot vote proposed Risk Steward roles with no delay for Aave V4. Their emergency powers work in one direction only: they can freeze or halt markets, but they cannot unfreeze them. CryptoSlate noted that the proposal requires no public report when those powers are used.

Compound. Compound's Community Multisig can pause the protocol, whitelist who may submit proposals, set borrow caps and, through the Proposal Guardian, cancel malicious proposals. Its current mandate runs from October 2025 to December 2026 with 7 signers, two of whom are the audit firms ChainSecurity and Certora. The sources reviewed here do not state its signing threshold. That gap is exactly what the second question on the checklist is meant to catch.

The strongest objection is that small key sets are what actually protect depositors, and the recent record supports it. The Kelp DAO bridge exploit drained $292 million, and Arbitrum's nine signatures recovered roughly a quarter of it. Term Finance, by contrast, had a seven-day timelock and a veto available to liquidity providers. No one used the veto, and an attacker who started with about 2 ETH removed $8.5 million through a vote the rules allowed. Patka's claim that over 90% of recent incidents come from operational or parameter failures points the same way: code is rarely the weak point.

None of this weakens the checklist. It strengthens it. The research reviewed here does not attribute a recent loss to a stolen admin key: Kelp's loss came from a single-verifier bridge setting, and Term Finance's came through governance. But a key that can save funds is, by construction, a key that can move them. The checklist does not ask whether such power exists. It asks who holds it and whether anyone would notice in time.

A token vote is not the same as decentralized control. Across 48 Ethereum DAOs studied in 2026, the top ten holders controlled more than half of the voting power in 39 of them. A multisig is also not the same as a DAO, even when a DAO elected it. The distinction between the code and the body that changes it is set out in the wider guide to protocols.

Run the six questions and write the answers down. Where the documentation is silent on the threshold, the signers or the delay, treat the protocol as controlled by whoever holds the admin role. The figure that matters most is the delay on upgrades compared with the time it would take a depositor to exit. If the delay is shorter, the protocol's decentralization does not protect that depositor.