From Mt.Gox to FTX: Shared Warning Signs Before 8 Exchange and Lending-Platform Failures

Eight cases read together

Mt.Gox (2014), QuadrigaCX (2019), FTX (2022): each of them, when it first happened, was easy to treat as an isolated event.

Read together, what was missing before each of the eight cases went wrong sorts into five categories: reserve accounts with no independent verification, operation in a single jurisdiction, no user-protection mechanism, affiliated trading firms not kept separate, and public statements from the people in charge that offered assurances but no mechanisms.

This lesson first takes the causes apart case by case (one short paragraph each), then pulls out the shared signals, and finally writes them up as five checks you can run against any exchange you use today.

1. The eight cases one by one

Mt.Gox 2014

Scale: at peak, over 70% of global BTC trading volume. Loss: the company announced roughly 850,000 BTC missing, and later found about 200,000. Root cause: according to Wizsec's 2017 on-chain investigation, the hot-wallet private keys were stolen in September 2011 and about 630,000 BTC were taken out over the following two years; when the company became aware of it has never been established. Early signals: in May 2013 US Homeland Security seized the accounts of Mt.Gox's US subsidiary; from June 2013 USD withdrawals were suspended, and complaints about delayed withdrawals kept coming. Full account in Shipwreck Annals · Vol. I.

Bitfinex 2016

Scale: one of the main USD/BTC venues at the time. Loss: approximately 120,000 BTC — about $72 million in 2016 dollars. Root cause: a multisig wallet architecture implemented with BitGo was compromised. Outcome: Bitfinex did not go bankrupt. All user balances were written down by about 36% and compensated with a "BFX token"; the exchange is still operating today. Afterwards: in 2022 the US Department of Justice announced the seizure of about 94,000 BTC linked to the case.

BitGrail 2018

Scale: small Italian exchange, but the primary trading venue for NANO. Loss: about 17 million NANO, roughly $170 million at the time. What happened: on February 9, 2018 BitGrail announced it had been hacked, and user withdrawals stopped. Outcome: victims brought a class action against founder Francesco Firano; in January 2019 the Florence court found that BitGrail had failed both to put safeguards in place and to report its losses (Wikipedia · Nano).

Cryptopia 2019

Scale: mid-sized New Zealand exchange. Loss: its hot wallet was breached in January 2019; the blockchain firm Elementus estimated the stolen assets at about NZ$24 million at the time, and the theft carried on for weeks afterwards (Wikipedia · Cryptopia). Outcome: it went into liquidation in May 2019.

QuadrigaCX 2019

Scale: Canada's largest crypto exchange at the time. Loss: according to the review published by staff of the Ontario Securities Commission (OSC) in June 2020, more than 76,000 clients were owed a combined CAD$215 million in assets and ultimately lost at least CAD$169 million. What happened: on January 14, 2019 Quadriga announced that founder Gerald Cotten had died in India the previous month; on February 5 the platform ceased operations and filed for creditor protection. Root cause: OSC staff's assessment is that most of the shortfall resulted from Cotten's fraudulent conduct — he opened accounts on his own platform under aliases, credited them with balances that did not exist, traded those against real clients, and covered the resulting losses with other clients' deposits (OSC report). These are the conclusions of regulatory staff and have not been tested before a court. Full account in Shipwreck Annals · Vol. IV.

Celsius 2022

What it was: a US crypto lending platform, not an exchange. What happened: it paused withdrawals on June 13, 2022 and filed for bankruptcy in July. Root cause: the US Federal Trade Commission's July 2023 complaint alleges that Celsius duped consumers into transferring crypto onto the platform and misappropriated more than $4 billion of user deposits (FTC announcement). Outcome: CEO Alex Mashinsky was indicted on seven federal counts including securities fraud; in December 2024 he pleaded guilty (to one count of commodities fraud and one of securities fraud), and in May 2025 he was sentenced to 12 years in prison (Wikipedia · Alex Mashinsky).

Voyager 2022

What it was: a crypto broker and lending platform, once listed on the Toronto Stock Exchange. Root cause: Voyager had lent approximately $670 million to a single counterparty, Three Arrows Capital (3AC); when 3AC failed in June 2022, the loan could not be recovered (Wikipedia · Voyager Digital). What happened: it filed for bankruptcy in July 2022.

FTX 2022

Scale: one of the largest global crypto exchanges before it failed. Loss: approximately $8 billion in customer-fund shortfall. Root cause: its affiliate Alameda Research used FTX customer funds. Full reconstruction is in Shipwreck Annals · Vol. III. Early signal: on November 2, 2022 CoinDesk reported how much of Alameda's balance sheet was FTT; on November 7 SBF was still tweeting "FTX is fine. Assets are fine." Four days later, FTX filed for bankruptcy.

2. The shared early signals

By root cause the eight cases fall into different groups — fraud and misappropriation by insiders, counterparty failures, wallet and security failures. Set side by side, though, what went wrong sorts onto five dimensions: whether reserves were independently verified, whether the venue operated in a single jurisdiction, whether there was a user-protection fund, whether affiliated trading or lending entities were segregated, and whether the public statements of the people in charge contained anything that could be checked. Bitfinex 2016 is the outlier — it went wrong for technical reasons rather than governance ones. These five dimensions, inverted, are the five checks.

3. The five sieves

Each check covers three things: why it matters, how to verify it, and what can be traced in public sources about a few major exchanges on that point. The scoring is yours to do, following the verification steps.

Sieve 1 · Proof of Reserves that exists at all: published monthly, covering the user liability side, and checkable by you

Why it matters. Without PoR, you are letting the exchange say "we have $X" without letting you check. FTX, QuadrigaCX, and Celsius never published a proof of reserves. PoR is first of all not a technical question; it is a question of whether the exchange is willing to let outsiders look at the books.

Where it came from. For the history of PoR, and why Mazars paused this kind of report for crypto clients in December 2022, see What Is Proof of Reserves (PoR)?; it is not repeated here.

Why the liability-side proof matters. With a Merkle Tree alone, you can check that your own balance is in the tree, but you cannot verify the total — in theory an exchange could slip in a few fake accounts with negative balances and push the total liability down. A zero-knowledge proof such as a zk-SNARK can show, without exposing any individual balance, that no account balance is negative and that all balances add up to the published total liability.

How to verify (step by step).

1. Open the exchange's official site, search "Proof of Reserves" or "PoR" — if the page does not exist, this sieve scores zero for that exchange.
2. Check who issued it. A report from an independent third party is more credible; where there is none, all you can do is check the on-chain addresses yourself and use the page's own lookup for your account.
3. Check the frequency. Monthly beats quarterly, quarterly beats annual, no frequency commitment means no PoR. A snapshot that is too old tells you nothing.
4. Check whether there is a verifiable proof on the liability side. An asset-only PoR can be padded with coins borrowed for snapshot day. If the report says "our BTC wallet balance is X" but does not say "what we owe users is Y, and X ≥ Y," discount it accordingly.
5. Check whether you can self-verify. A good PoR provides a lookup that lets you confirm, with your own account identifier, that your balance is included in the most recent snapshot. Whether a given exchange offers one, check its own PoR page.

How the exchanges compare on this sieve: Binance, OKX and Kraken all have official PoR pages (linked in the references at the end); Coinbase is a listed company and takes the route of regular financial disclosure. For every one of them, walk through the steps above yourself, and read each venue's current reserve ratios on its own PoR page, which changes all the time.

Sieve 2 · At least three meaningful financial licences across major jurisdictions

Why it matters. Operating in a single jurisdiction puts all the policy and regulatory risk in one place: Cryptopia operated only in New Zealand, BitGrail only in Italy. Being licensed in several jurisdictions means more than one regulator has looked at the compliance material, and if something goes wrong in any one of them, there is one more place a signal can come from.

Which licences carry more weight. There is no agreed ranking of how hard each jurisdiction's licence is to get. Go by the licensee register and the scope of permission on the regulator's own site, not by how the exchange describes it.

How to verify.

1. Open the exchange's "Regulation" or "Licenses" page — usually in the footer or About sub-page. If the page does not exist, or only displays an image without licence numbers, score zero.
2. Verify each licence at the regulator's public registry. FCA, VARA, FINTRAC, FSA, NMLS, MAS, SFC, BaFin all run public licensee directories; the company name and licence number are enough to confirm.
3. "Applied for," "in discussion," "in-principle approval" do not count — only licences actually issued count.
4. Verify the licence is still in force. Go by the current status in the regulator's register; what is displayed on the exchange's own site does not count.

For comparison · examples of licences for the main exchanges that can be traced in public sources (checked September 2026; not a complete list).

This table lists only items that can be traced to a source. It is not a complete licence list for any of these exchanges, and it is not a ranking by count — each of them holds more licences than shown here, and the picture changes every year. To judge an exchange, follow the steps above and check each licence in the regulator's own directory — only what shows as "issued" counts.

How the exchanges compare on this sieve: each exchange's licensing changes every year, and the table above is only a set of examples. Whether the exchange you intend to use holds a formal licence in your own jurisdiction is the thing on this sieve most worth checking in the register yourself.

Sieve 3 · Public user-protection fund (with public dollar amount and on-chain verifiable wallet)

Why it matters. When an exchange is hacked or fails, if no money was set aside in advance in a form outsiders can check, users simply join the bankruptcy queue — that is what happened after the Cryptopia hack. A user-protection fund is meant for exactly this situation.

Three structural forms of "insurance fund" — they are not equivalent.

• On-chain segregated wallet (for example Binance SAFU): the exchange keeps the fund at a publicly published on-chain address, and the balance is checkable by anyone. Note that whether such a fund pays out, and how much, is the exchange's own decision, not a legally mandated obligation (see SAFU explained).
• Balance-sheet reserve: the exchange carries a line on its corporate balance sheet labelled "risk reserve" but does not segregate it into a separate wallet, so it cannot be checked on chain. This is still the exchange's money — if the exchange goes bankrupt, it is pulled into the bankruptcy estate.
• Commercial insurance (for example Coinbase): the exchange buys a commercial policy from an insurance company. Strength: there is a right to claim. Weakness: policy terms are usually not public, coverage caps are limited, and policies usually cover only specific events (such as hacks), not operator misappropriation.

Side-by-side fund comparison.

• Binance SAFU: set up in July 2018, when Binance announced it would route 10% of trading fees into a separate wallet. The language changed later on: the public position became that the company monitors the size of SAFU and keeps it at a level sufficient to protect users, rather than skimming a fixed share off every trade. On 18 April 2024 Binance published "Update: SAFU Conversion to USDC", converting the fund entirely into USDC and describing its size as something that fluctuates, though "this level is typically set at $1B". That post named one USDC address, 0x420ef1f25563593aF5FE3f9b9d3bC56a8bd8c104; Binance's official SAFU FAQ (updated 2026-02-02) now publishes two addresses — that USDC address plus the bitcoin address 1BAuq7Vho2CEkVkUxbfU26LhwQjbCmWQkD — so the balance is something any reader can verify on-chain instead of taking on trust. Exactly one draw-down has been publicly confirmed: in May 2019 Binance itself lost roughly 7,000 BTC to attackers (around $40M at the prices of the day) and announced that SAFU would absorb all of it. The 2020 KuCoin theft and the 2022 BSC Token Hub exploit get filed as SAFU draw-downs all the time, but neither one is, going by the official record — the long version is in SAFU explained.
• Coinbase: takes the commercial-insurance route, and has insured the bitcoin held on its own servers since 2014 (Wikipedia · Coinbase). The amount of cover and the policy terms have not been published in full, and commercial policies usually pay out only for specific events (such as hacks), not for operator misappropriation; nor can this kind of insurance be verified on chain.
• OKX: back in 2018, after a trading incident, it put 2,500 BTC into an insurance fund (Wikipedia · OKX).
• Bybit: in February 2025 it lost about 400,000 ETH, worth about $1.5 billion by the FBI's figure, in a hack the FBI attributed to North Korean hackers. Bybit said it replenished its reserves within 72 hours by securing about 447,000 ETH through emergency funding (Wikipedia · Bybit). The hole in this case was filled with outside funding, not from an insurance fund set aside in advance.

How to verify.

1. Look for either a public dollar amount or a public on-chain wallet address with a verifiable balance. Neither = score zero.
2. Check the structural form — balance-sheet reserve, or a segregated trust / on-chain wallet. A segregated account at least lets outsiders check the balance.
3. Calculate the ratio. Divide the insurance fund by the total user assets shown on the exchange's PoR page to get a rough proportion; both numbers change every period, so use the period you checked. An insurance fund whose share is small enough to ignore is more symbolic than substantive.
4. Check whether the fund has ever actually been drawn down — and insist that every draw-down you count has an official source behind it. For Binance SAFU the only publicly confirmed use is the May 2019 theft from Binance's own hot wallet. The draw-down lists circulating online mix in events such as KuCoin 2020 and BSC Token Hub 2022, where SAFU was not touched at all, so copying those lists inflates the record. A fund never drawn down, or one whose draw-downs you cannot trace to an official announcement, has unproven operational mechanics.
5. Check disclosure cadence. One-time announcement followed by silence is a deduction.

How the exchanges compare on this sieve: go through the five steps above one by one, and count anything you cannot find as "no."

Sieve 4 · No related-party trading firm, or independent audit of segregation

Why it matters. FTX and Alameda, and Voyager with its single counterparty 3AC, are both cases of related-party or counterparty risk. If an exchange also owns a market maker, or depends heavily on a single counterparty, it should be able to produce a report from an independent third party showing the two are separated. FTX's problem was structural: SBF controlled both the exchange FTX and the trading firm Alameda, and FTX customer funds were lent to Alameda. The same move at a regulated broker would simply be misappropriation of client assets.

Counterexample · FTX and Alameda. When FTX was founded in 2019, SBF was already running Alameda; in 2021 FTX moved its headquarters from Hong Kong to the Bahamas (Bloomberg report). Nominally, two companies. But FTX co-founder Gary Wang testified that in 2019, at SBF's direction, he changed the code so that Alameda could draw nearly unlimited funds from FTX without triggering automatic liquidation (Wikipedia · United States v. Bankman-Fried). The term for this structure is "commingling": one person controls Company A and Company B, and when Company B loses money, Company A's customer funds can flow to Company B.

For comparison · a listed company's disclosure duty. Coinbase listed directly on NASDAQ in April 2021. A listed company has to disclose related-party transactions in its periodic reports, which is a layer of outside constraint imposed by law. It does not mean nothing can go wrong, but it makes problems easier to see.

For comparison · public facts that can be traced about the main exchanges on related parties (checked September 2026).

How to verify.

1. Search whether the controlling shareholder or CEO simultaneously controls other crypto companies. OpenCorporates, SEC EDGAR, Hong Kong Companies Registry all free.
2. If affiliated entities exist, check whether SOC 2 Type 2 or equivalent independent audit attests to segregation. SOC 2 Type 2 is an AICPA standard that specifically audits "control effectiveness"; if an exchange says it has one, ask for the report or for the statement on its official site.
3. Check whether the exchange discloses its largest single-counterparty exposure (Voyager didn't; FTX didn't disclose Alameda's share). An exchange that discloses this voluntarily earns extra credit.
4. Check CEO LinkedIn and corporate-registry filings — a CEO concurrently serving as director of multiple crypto entities loses 0.5 of a point.
5. Platform tokens. An exchange issuing a token is not itself a problem, but the token being pledged, used as collateral or posted as margin by the exchange or its affiliates is a related-party risk. FTX's FTT is the example.

How the exchanges compare on this sieve: listed companies have a legal duty to disclose related parties; for unlisted exchanges, you can only check their shareholders, affiliated entities and platform-token uses yourself, following the five steps above.

Sieve 5 · CEO public statement track record

Why it matters. The test on this sieve is whether he answers a risk question with a mechanism or with an adjective: how often the books are reconciled, who runs the third-party audit, how frequently reserves are attested, who makes the call when something breaks — those are mechanisms. "We take risk extremely seriously," "security is our top priority" — those are adjectives, and an adjective needs nothing behind it. Anyone can produce one. This sieve carries the same weight as the other four not because it is more accurate, but because it is the cheapest to run: the other four mean reading filings, this one means listening to two podcasts.

One example. On November 7, 2022, SBF tweeted "FTX is fine. Assets are fine." (Wikipedia · United States v. Bankman-Fried); four days later FTX filed for bankruptcy. The sentence contains an assurance and nothing that can be checked.

Where to find a record you can compare against. A listed company's quarterly earnings calls are a ready-made archive — how the same person answers questions about reserves and risk from one quarter to the next can be compared. For unlisted venues, go to interviews, podcasts and past statements.

How to verify.

1. Search the CEO's name plus "podcast" plus "interview" — listen to 2-3 of the most recent public conversations. YouTube and Apple Podcasts are the direct channels.
2. Watch for four negative signals: (1) risk-appetite bragging ("we're the most aggressive"); (2) regulator-contempt remarks ("regulators don't understand"); (3) ad-hominem on competitors; (4) blanket dismissal of every concern as "FUD."
3. Watch for three positive signals: (1) willing to acknowledge errors using specific words rather than "we are constantly improving" boilerplate; (2) cites specific numbers (user count, reserve ratio, liquidity depth); (3) discusses risk-management details (team size, tools, audit cadence).
4. Look longitudinally across five years — a CEO's consistency over five years is more informative than any single interview. Twitter history plus older interview videos are all free archives.
5. Confirm whether the CEO personally appears on quarterly earnings calls (public) or monthly AMA (private). A CEO who lets CFO/PR substitute the entire call loses 0.5 of a point.

How the exchanges compare on this sieve: this one you can only listen for yourself. A listed-company CEO has the quarterly earnings calls as a ready-made record; for the rest, go through their interviews and past statements using the five steps above.

4. How to score, and what the score means for position size

The method is simple: a sieve passed scores 1, half passed scores 0.5, failed scores 0, for a maximum of 5 across all five. Please do the scoring yourself — every sieve above comes with verification steps. This lesson gives no verdict on any exchange.

Decision tree · How scores map to position sizing

Having the score is only half the work — you also need a rule that maps it to a position size. The rule below cuts the 0-5 range into four tiers. It is one way of splitting; how much to hold in each tier is capped by the "kept on an exchange" share that Lesson One gives for your amount tier:

• Main tier (score ≥ 4.5/5): can serve as your main exchange position. Which exchange makes this tier is decided by the score you give it yourself.
• Secondary tier (3.5/5 ≤ score < 4.5/5): a small share only, as a backup. Its uses: (1) evacuation channel if your primary pauses withdrawals; (2) backup for assets or regions your primary does not support. A secondary venue should not carry long-term holdings.
• Cautious tier (2.5/5 ≤ score < 3.5/5): small experimental amounts only. The defining characteristic of this tier is "strong on some dimensions, clearly weak on others" — not for long-term funds.
• No-tier (score < 2.5/5): zero allocation. This is not a claim that such an exchange will fail; it is a statement that it lags clearly on the structural sieves and is not worth the risk.

How often to redo it: redo the scoring every six months. Redo it straight away if something significant happens: the exchange you use receives a regulatory penalty, the people in charge say something alarming in public, or a PoR period is missed.

Two notes on scoring.

1. The score covers structural risk only. When two exchanges score close together, the remaining differences are product features, regional availability, asset support and other secondary dimensions — not covered in this lesson.
2. The score reflects only the moment you scored. Each venue's situation changes, and the five checks themselves will change with the industry.

5. The limits of this framework

Limit 1 · Hidden misuse of funds cannot be seen from outside

Before FTX failed, what outsiders could see was that it had licences, celebrity endorsers and large institutional investors. It had no proof of reserves, which costs points under this lesson's checks; but the fatal problem — customer funds being used by its affiliate Alameda — only came out afterwards, through the trial and the bankruptcy proceedings (see Shipwreck Annals · Vol. III). An outside user could not have seen it in October 2022. Hidden misuse of funds between affiliated companies is invisible from outside, and that is this framework's fundamental blind spot.

So passing all five checks does not mean safe. The framework screens out structural risks that can be seen and is nearly powerless against hidden internal misuse. Even with a main-tier exchange, Lesson One's advice still applies — large, long-term funds belong in self-custody, and an exchange holds only what you intend to trade in the near term.

Limit 2 · In 2014 these checks did not yet exist

Looking back at Shipwreck Annals · Vol. I: when Mt.Gox failed, it met hardly any of the five — no proof of reserves, no insurance fund, and its US accounts had been seized by a regulator. But nobody in 2014 was measuring exchanges against these five checks; proof of reserves that also covers liabilities only took shape after Mt.Gox went down. Grading past events by today's standards makes the standards look effective, but they had not taken shape at the time. In the same way, today's five may not be enough a few years from now, so the framework needs revising periodically and should not be treated as fixed.

Limit 3 · The scoring was fine — the problem was position size

Measured against the five checks, Celsius passes very few. It had no proof of reserves; it publicly claimed a $750 million deposit insurance policy, and the FTC's 2023 complaint alleges that it held no such policy and that it misappropriated more than $4 billion of user deposits (FTC announcement). Under the decision tree, a platform like that lands in no tier. So the problem Celsius exposes is not in the scoring but in whether people sized their positions by the score: the framework only helps people who actually follow the tiers. If a no-tier platform offers a very high yield and you put most of your funds there anyway, the framework cannot help.

Limit 4 · Things these five checks do not cover

• Off-balance-sheet business. Off-book movement of funds is not something an ordinary user can verify, and the five checks do not reach it.
• Withdrawal delays. Delays tend to show up before withdrawals are paused, as they did at Mt.Gox. They are not a separate check, but they are worth watching for at all times: reduce your position as soon as they appear.
• Anti-money-laundering enforcement history. On November 21, 2023 Binance pleaded guilty to charges including violations of the US Bank Secrecy Act and settled with the US Treasury's FinCEN and OFAC, with penalties totalling about $4.3 billion (US Treasury announcement; Wikipedia · Binance). This kind of record is not a separate check. What matters when you look at it is whether similar violations have recurred since — check the latest regulatory announcements yourself when you score.