Markets open ยท Independent crypto analysis November 8, 2026
Markets

Custodial Failures Are Rarer, Bigger, and Far More Likely to Be Repaid

Exchange failures are 6.6% of recorded incidents and 26.6% of the money — and 37.9% of what they lose is later returned. For DeFi protocols the recovery rate is 8.5%.

The CoinageReport Desk's avatar
The CoinageReport Desk
ยท 4 min read
X
A blank notebook and pen, representing the responsibility of writing down and safeguarding a private key or seed phrase

Sort DefiLlama’s 610-entry theft register by the type of thing that was attacked and the custody argument changes shape. Centralized exchanges account for 6.6% of incidents and 26.6% of the money.1 DeFi protocols account for 81.5% of incidents and 46.9% of the money. The median exchange incident is $41.50 million; the median protocol incident is $2.01 million. Pulled 3 August 2026.

Then there is the figure nobody puts on a slide. Of the value taken from exchanges, 37.9% was subsequently returned. From DeFi protocols, 8.5%. From the nine incidents classified as wallet attacks, nothing at all.

What that does to the slogan

“Not your keys, not your coins” is a true statement about control. It is routinely used as a statement about expected loss, and on this register it does not survive that use. Moving assets out of custody moves you out of the only category with a recovery rate near two-fifths and into categories where recovery runs at 8.5% and zero. That is a real, measurable cost of self-custody, and the slogan is constructed to hide it. Our position is that anyone repeating the phrase as financial reasoning rather than as a description of who can freeze your account should be asked for their recovery numbers.

The opposite case is also in the data, and it is not weak. Exchange failures are rare and enormous. Since the start of 2025 there have been 11 exchange incidents taking $1,866 million against 198 protocol incidents taking $1,621 million.1 Eighteen times fewer events, more money. If your entire position sits in one custodial account, the register says your tail risk is a single event, and that event has historically been large enough to matter.

The register, by target

  • DeFi protocol — 497 incidents — $7,931m — median $2.01m — 8.5% returned
  • Exchange — 40 incidents — $4,495m — median $41.50m — 37.9% returned
  • Other — 22 incidents — $4,057m — median $6.70m — none returned
  • Wallet software — 9 incidents — $262m — median $7.00m — none returned

What this does not show

It does not measure self-custody. Nine wallet entries is far too small a sample to generalise from, and the dominant failure mode of holding your own keys does not appear in this dataset at all: lost seed phrases, approval phishing aimed at individuals, and holders who die without a recovery plan generate no post-mortem and no register entry. Self-custody’s real loss rate is unmeasured here, and it is not zero.

It also excludes insolvency. Exchanges that failed because they lent out customer assets were not hacked, so they are absent, and that materially understates the custodial column. The returned-funds field records value recovered, not customers made whole; a firm recovering coins is not the same as a firm reimbursing account holders, and reimbursements paid from a balance sheet are not counted either.

Most importantly, there is no denominator. We do not know how much value is held custodially versus non-custodially at any point in the period, so none of these totals can be turned into a loss rate. Absolute dollars are the only comparison available, and absolute dollars favor whichever side holds less. We will re-run this split on 3 November 2026. If a credible published series for custodial versus self-custodied assets under management appears, we will use it and recompute everything as rates.

How self-custody actually works

This note compares two failure modes and their outcomes. The mechanics of the non-custodial side — seed phrases, backups, the threat model in order of how often each part fails, and how to check a hardware vendor rather than trust one — are set out in our standing reference on self-custody and wallets.

Sources
  1. DefiLlama, hacks register, grouped by target type. 610 entries, 17 June 2016 to 2 August 2026. Pulled 3 August 2026. ↩ ↩

Discover more from CoinageReport

Subscribe to get the latest posts sent to your email.

This article is for informational purposes only and is not financial advice.

Get the next report in your inbox

Weekly market notes and analysis. No noise.

TH
Written by
The CoinageReport Desk

An editorial byline, not a pen name. Pieces published under the Desk were researched, their figures independently re-checked against source, and reviewed before publication. Editorial responsibility rests with the Editor in Chief.