808bits

The DORA exit plan for custody is usually one untested sentence

2026-06-05 · updated 2026-11-10 · 5 min · custody operations

Most custody exit plans I have read give the exit itself one sentence: “In the event of provider failure, keys will be recovered from the encrypted backup using the provider’s published recovery tooling.” It is true as far as it goes. The most common gap in the plans that carry it is a date: the last time anyone proved the backup works was three years ago, or there was no last time.

That sentence has had a regulation pointed at it since 17 January 2025. DORA, the EU’s Digital Operational Resilience Act, says in Article 28(8) that every financial entity must keep exit strategies for the ICT services supporting its critical or important functions, and that the plans “shall be comprehensive, documented and, in accordance with the criteria set out in Article 4(2), shall be sufficiently tested and reviewed periodically”. The technical standard beneath it, Commission Delegated Regulation 2024/1773, adds in its Article 10 that each plan must be “realistic, feasible, based on plausible scenarios and reasonable assumptions” and carry “a planned implementation schedule”. Documented is the easy word. The sentence above cannot show tested or feasible.

I read both texts and mapped every requirement onto what an exit from a custody platform has to contain. What the mapping taught me is that custody is the one ICT service where the regulation’s own frame, remove the service and the relevant data, points the wrong way. What has to come out of the vendor is not data. It is the ability to sign.

On a multi-party computation platform, MPC for short, the private key never exists in one place. It lives as shares, some held by the vendor and some by the customer, and signatures are computed between them without assembling it. The customer’s half of the exit is a recovery backup, an encrypted file cut on one day plus a recovery key that someone keeps. The exit question is whether that file can be turned into signatures with the vendor’s systems switched off, and it splits in three. Can the keys be reconstructed from the customer-held backup alone? Does the backup cover every keyset the workspace has created since, or only what existed on the day it was cut? Once the keys are back, can everything move, including what is staked, locked, or sitting inside a smart contract?

A plan that answers none of those has described a hope.

The vendors answer them differently, and that is where the one-sentence plans fail first. I went through the public exits in the order I found them. Fireblocks publishes an open-source Recovery Utility that rebuilds extended keys from the customer’s backup kit on an offline machine. I have read its verify path, and it is a real exit with one boolean between verifying and recovering. Cobo publishes recovery kits on GitHub that reconstruct keys from exported recovery group files, which its manual says are held by the customer’s organisation or a trusted third party. BitGo’s wallet recovery wizard builds unsigned sweeps from the user and backup keys of its multisig wallets without BitGo’s service in the loop. Copper’s vaults split the key across three independent parties with a 2-of-3 signing quorum, so that client plus the third party can sign without Copper. Other vendors route recovery through a named backup provider rather than a tool the customer can run.

Those are five different exits, with five different kinds of evidence, and the plans I have read do not say which one applies to them. Where a vendor offers none of the five, the absence is itself a finding under Article 29, the concentration-risk article, and it belongs in the register of information, not in a footnote.

The objection I hear is that the vendor has this covered. Several now market DORA exit support as a feature, and it is better than nothing. It is also the one piece of evidence a reviewer discounts, because the scenario being planned for is the one where the vendor is gone. A vendor attesting to its own escapability is the plan writing its own reference.

What a reviewer wants instead is the same five things every time. A date on the last verification run, with the tool version pinned. A coverage number, keysets in the backup against keysets in the workspace, because backups are point-in-time and workspaces grow. Proof the verification ran without the vendor in the room. A timeline for the assets a sweep cannot move, in the days and weeks of unbonding that the regulation’s “implementation schedule” turns out to mean. And a rehearsal, because the text says tested. A recovery drill on a test workspace, run by the people who would do it for real, produces all five. Before the drill, a twelve-question self-assessment shows in about five minutes which of them are missing.

DORA binds EU financial entities, but the perimeter reaches further than it looks from Zug or Zurich. EU group entities are in scope directly. EU clients push the requirements down by contract. And FINMA, the Swiss supervisor, has asked Swiss banks the same questions under different numbering since 1 January 2024, through its Circular 2023/1 on operational risks and resilience. The five items satisfy both regimes. Only the letterhead of the reviewer changes.

None of the plans I have read contained all five. Most contained none, and the sentence at the top instead.

Correction. An earlier version cited a Copper page that does not describe the 2-of-3 arrangement, and wrote “critical functions” where the regulation says “critical or important”.