The Underwriting Crisis

Crypto's titans repeated these four mantras until the industry mistook them for first principles. All four are true. None of them was finished, and the unfinished half cost $55 billion.

Each was true enough to believe and incomplete enough to be lethal. Each skipped the same unglamorous thing: the continuous, painstaking work of underwriting risk. And the people repeating them loudest were the ones with the most to lose if anyone finished the thought: the mantras justified the valuations and held the moral high ground over the TradFi system the industry claimed to be replacing.

So the work never got done.

Crypto has an underwriting crisis. It has always had one.

3AC, Alameda, FTX, Genesis, Voyager, Celsius, Terra, BlockFi, Mango Markets, Ronin, Stream Finance, KelpDAO, Elixir, and Resolv are not outliers. They are the predictable output of an industry that scaled capital faster than it built the due diligence, accounting, and security infrastructure to govern it.

Since June 2021, conservatively more than $55 billion has been destroyed. The number is built from three overlapping buckets: roughly $40B erased in the span of a week when Terra unwound, plus the cascade of CeFi lenders and exchanges that died around it; on the order of $16B stolen through technical exploits between 2021 and early 2026; and the opening bill of a third era now underway, headlined by the $1.5B Bybit theft and the curator-driven contagions of late 2025. The eras overlap, the figures are approximate, and they are almost certainly conservative.

What unites every entry on that list is not a single failure mode. It is a single absent function. There was never a continuous underwriting layer sitting between capital and the infrastructure it depended on: no one systematically tracking solvency, counterparty exposure, and configurations, and pricing counterparty risk as conditions changed. The cumulative losses are what that absence amounts to at scale.

Continuous underwriting is not exotic. Traditional finance calls it the credit committee, the rating, the custodian, the auditor, the vendor risk desk, each one built over time in the wreckage of a specific disaster, because finance learned, repeatedly and at enormous cost, that capital deployed without continuous underwriting incinerates itself. Crypto skipped that century of hard lessons because of the speed of smart-contract experimentation, and is now relearning them in compressed time. The parallels in this piece are not decorative. Every failure crypto treats as unprecedented has a named precedent in markets that finance professionals lived through.

This thesis makes one argument, stated several different ways across the eras and failure modes:

Fixing crypto's structural underwriting problem starts with an examination of this core truth throughout its history: transparency alone guarantees neither security nor solvency.

In November 2022, FTX became the point of no return for thousands of companies and hundreds of thousands of users. An $8B hole in the balance sheet, first surfaced by CoinDesk, sat beneath a structure with no auditors worth the name, no functioning board oversight, commingled customer funds, and fractional reserves. When the gap became visible, the result was an ordinary bank run on an extraordinarily fragile institution.

None of these failure modes was new. Commingling customer money with the house's own book is precisely what destroyed MF Global in 2011, when Jon Corzine's firm dipped into roughly $1.6 billion of segregated customer funds and a former U.S. senator and Goldman CEO walked his brokerage into bankruptcy. A balance sheet that simply was not what it claimed is Enron in 2001 and Wirecard in 2020, the latter a DAX-listed company whose auditors signed off for years on €1.9 billion of cash that did not exist. And a run on a fragile-but-"solvent" institution is Silicon Valley Bank in March 2023, killed in roughly 48 hours by a digital deposit run before regulators could open on Monday.

The damage did not stop at FTX's own customers. An entire daisy chain of lenders had been extending credit to each other with little collateral and no shared visibility into counterparty exposure. That is a sample, not a census: an interlocking web of leveraged institutions, each invisible to the others' books, collapsing in sequence the moment one node failed. Finance has seen this film before. It is Long-Term Capital Management in 1998, whose opaque, hyper-leveraged positions were spread across more than a dozen banks that each thought they understood their own exposure and none of whom could see the whole. It is AIG in 2008, the counterparty hiding inside everyone's risk model at once.

Crypto's response to FTX was almost liturgical: this is why we need DeFi. Opaque TradFi systems broke; the transparency of public blockchains and the smart contracts running on them would restore trust in a way that centralized finance never could.

There is a fatal problem with that story. It's wrong. The mantra was too simple, and therefore optimistic for the wrong reasons. It puts the cart before the horse with respect to what transparency alone can accomplish in isolation, especially considering that one of DeFi's largest projects blew up six months before FTX.

That precursor was Terra, an economic engine operating entirely in plain sight that still vaporized roughly $40B in the span of a week in May 2022, most of it inside 48 hours, with every asset (LUNA) and liability (UST) visible to everyone the entire way down. The system was never actually held together by the solvency it advertised. It was held together by available liquidity for UST and LUNA across exchanges, and when that liquidity gave way, transparency did nothing to stop it.

Two TradFi failures live inside Terra simultaneously. The first is the failed currency peg defended with finite reserves: this is Black Wednesday in 1992, when the Bank of England burned through its reserves trying to hold sterling's ERM peg until George Soros and the market simply overwhelmed it. A peg holds until the reserves backing it are exhausted, and then it does not hold at all. The second is breaking the buck, the Reserve Primary Fund in September 2008, a money-market fund that everyone treated as a dollar until its Lehman exposure pushed its share price below $1.00 and triggered an industry-wide run. The lesson, in finance's own vocabulary, was an asset-liability-matching lesson: the volatility and liquidity profile of the underlying assets is what determines whether a structure survives stress, and this was knowable in advance.

Eleven months before Terra, in June 2021, Iron Finance's TITAN had already run the same play at smaller scale: a partially algorithmic stablecoin backed in part by stables and in part by a volatile reflexive token, spiraling to zero in a textbook death spiral that almost nobody remembers.

Two collapses, both fully transparent, both modelable, both forgotten in the name of a more positive mantra, "DeFi fixes [insert TradFi break]", and both occurred before FTX. That timing is why they got memory-holed. The post-FTX narrative needed CeFi to be the villain and DeFi to be the cure, and Terra and Iron Finance were inconvenient counter-evidence that transparent, onchain systems fail for exactly the same underwriting reasons opaque ones do.

The most devoted defenders reframed them as "experiments gone wrong." They were not experiments gone wrong. They were systems permitted to grow far past the point their asset-liability mechanics could support, because nobody was properly modeling and surfacing those mechanics.

Venture capital amplified every layer of this. FTX raised nearly $2 billion across successive rounds from investors including Sequoia, Paradigm, SoftBank, and BlackRock. Terraform Labs raised over $200 million from Galaxy Digital, Pantera, Coinbase Ventures, and Lightspeed, several of whom also backed FTX. The capital structures were circular in ways that should have been disqualifying: Bankman-Fried quietly routed more than $500 million back into funds run by the same firms writing him checks. Related-party financing that loops capital back to its source to manufacture the appearance of validation is, again, not new. It is the engine of Enron's off-balance-sheet SPEs and Wirecard's round-tripped revenue. And none of the crypto VC firms just mentioned (among the most sophisticated allocators in the world, with full access to data rooms and management) caught an $8 billion hole or an algorithmic stablecoin with no survivable stress scenario.

The uncomfortable explanation is structural, not personal. The demand for fast exposure to speculative retail flow turned diligence into a game of commit now or lose the round, a pressure that projects with perceived momentum learned to weaponize. Even blue-chip investors failed to resist it, the same FOMO that drove sophisticated money into pets.com in 1999 and into AAA-rated subprime paper in 2006.

By January 2023, when Genesis filed and the last Contagion-era domino fell, the industry had its story straight: CeFi was the disease, DeFi was the cure. It was comforting. It was also wrong, proven wrong twice before FTX ever collapsed. Public data is not equivalent to visualized, understood risk. The Exploit Era was about to teach the same lesson in a language the industry could not wave away.

If the Contagion Era's article of faith was that transparency would restore the trust CeFi had betrayed, the Exploit Era ran on a second creed, narrower, more operational, and far more defensible: "Don't trust, verify."

Crypto inherited it from the open-source movement that birthed it, and underneath it sat one of software's most-quoted axioms, Linus's Law, coined by Eric Raymond in 1999: given enough eyeballs, all bugs are shallow. Open the code. Invite the world to read it. Pay a name-brand firm to audit it. Transparency was not just a virtue; it was the security model.

And the creed is not wrong. Don't trust, verify is one of the few genuinely load-bearing ideas crypto produced. The problem was never the creed. It was what the industry quietly did with it. It heard verify (an activity, something you do, continuously) and it built verifiable (a property, something a system has, once). Then it treated the second as if it were the first. Open the code, and it counted as checked. Publish the ledger, and it counted as watched. Pass a name-brand audit, and it counted as safe. A quiet conversion, from verifying to having-been-verified, and almost nobody noticed it happen.

Worse, a kind of passive momentum grew on top of it. The longer a contract sat in the open without being drained, the safer it was assumed to be, as if elapsed time in public were itself a form of diligence. Call it the lindy fallacy of security: survivorship mistaken for soundness, visibility mistaken for vigilance. Time, the story went, was on our side. It was not.

Software and finance both already knew this, and the proof predates crypto's worst exploits. In April 2014, the world learned about Heartbleed, a catastrophic flaw in OpenSSL, the single most-scrutinized piece of security code in existence, open and ubiquitous and reviewed for years. And the bug had been sitting in public, in plain view, for roughly two years anyway. Enough eyeballs to fill a stadium, and the bug was not shallow. Linus's Law is not a law. It is a hope, and Heartbleed was the moment the hope failed in the one codebase that should have been immune.

That is the lesson the Exploit Era taught crypto over and over: auditable code is not secure code. Verifiable is not verified. Transparency tells you the code can be read; it does not tell you anyone read the part that mattered, or that the part that broke was even in the code at all.

On March 23, 2022, $625M was stolen from the Ronin bridge. The Ronin team did not notice. For six days the funds sat in an attacker's wallet while the bridge kept processing deposits from users stacking fresh money on top of an emptied vault. The breach surfaced on March 29, not from an alarm, not from a monitoring system, but from a user who could not withdraw 5,000 ETH.

The vulnerability was not hidden. In November 2021, during heavy network congestion, Axie DAO had whitelisted Sky Mavis to co-sign transactions on its behalf. The arrangement was wound down that December, but the whitelist access was never revoked, leaving Sky Mavis with signing authority over enough validator keys to drain the bridge alone: nine validators, five signatures required, four already controlled by one entity. The concentration was structural, onchain, and public. North Korea's Lazarus Group only needed one more.

That structure, one party holding enough control to move the money with no independent check, is the oldest operational-risk failure in banking. It is Barings in 1995, where Nick Leeson held both the trading and the settlement function and used that absence of segregation of duties to hide losses that destroyed a 233-year-old institution. It is Société Générale in 2008, where Jérôme Kerviel ran €4.9 billion of unauthorized exposure using his knowledge of the bank's own back-office controls. Ronin's validator concentration would not have survived the first hour of a bank operational-risk review.

And then there is the case that turns the entire creed inside out. In August 2022, a routine upgrade to the Nomad bridge set its trusted root to 0x00, a change that made every withdrawal message valid by default, in public, auditable, open-source code. The "many eyes" were there. They found the bug almost immediately. And then they used it. The exploit required no skill whatsoever: copy a working transaction, swap in your own address, rebroadcast. More than 300 addresses piled in to loot $190M in what became DeFi's first crowd-sourced bank robbery.

You cannot tell the story of the Exploit Era without telling the story of Lazarus. By early 2026, the cumulative total attributed to DPRK actors across roughly 270 documented incidents sits near $6.75B, with proceeds funding North Korea's weapons program. In 2025 alone, DPRK-linked actors stole $2.02 billion, a 51% year-over-year increase; through April 2026 they accounted for roughly three-quarters of all crypto hack value worldwide. What makes Lazarus matter to an underwriting thesis is not the scale. It is the evolution. Their attack vector moved in lockstep with wherever the industry placed its trust, and in every case the information needed to anticipate the breach existed beforehand.

At Bybit, in February 2025, Lazarus compromised the development infrastructure behind Safe{Wallet} and swapped a legitimate JavaScript file for a malicious one that detected Bybit-specific addresses and rewrote transaction parameters on the fly. When the cold-wallet team initiated what they believed was a routine transfer, the interface displayed the correct details while the actual onchain transaction routed to Lazarus. The signers did everything right. They reviewed the transaction. They approved it. They were looking at a lie.

This is the deepest TradFi parallel in the entire era, and it is Madoff. Madoff's investors, their auditors, and the feeder funds that funneled tens of billions to him all reviewed statements, all signed off, all approved, because every one of them was looking at fabricated documents. It is Wirecard, where auditors relied on confirmation letters for cash balances that were simply forged. A control is only as good as the integrity of the information feeding it, and when the input itself is a lie, the most rigorous approval process in the world ratifies the fraud. You cannot verify your way out of a compromised input.

For scale: JPMorgan spends well over $600 million a year and fields thousands of dedicated security staff on cybersecurity alone. Tellingly, it also runs cybersecurity due diligence on counterparties before extending credit, because a borrower's security posture is a credit input. DeFi protocols securing comparable value spend a rounding fraction of that and conduct no such review of the systems they lend into.

Lazarus was the most sophisticated adversary in the era, but sophistication was rarely the requirement. The pattern across the largest failures is not attacker brilliance. It is that nobody was watching a live system that was already visible, and the watching, not the auditing, is the part crypto skipped.

TradFi built its monument to this exact lesson on August 1, 2012, when Knight Capital deployed new trading code and inadvertently reactivated dormant logic on one server. The result was a torrent of erroneous orders into the live market, with no circuit breaker to halt it. In roughly 45 minutes, Knight lost about $440M and nearly destroyed itself. The code had been written, reviewed, and shipped by professionals. What did not exist was a system watching the deployment behave in production and a switch to kill it when it misbehaved. An audit is a verdict on the code at rest. Knight died from the code in motion, and so did almost every protocol below.

Mango Markets deserves a second look, because it is the cleanest crypto instance of a crime TradFi spent decades building infrastructure to stop. The attacker manipulated a thinly traded asset's price to extract value against it, the same maneuver as the Hunt brothers' 1980 corner of the silver market, or the spoofing behind the 2010 Flash Crash, or the LIBOR and FX benchmark riggings of the 2010s. Traditional markets run mandatory, real-time trade surveillance (FINRA, exchange desks, the SEC's MIDAS system) precisely because manipulating a manipulable price is the most predictable attack in finance. Mango had no surveillance at all. A TradFi system would have flagged the 2,300% move in seconds.

The Exploit Era proved that auditable code is not secure code, exactly as the Contagion Era proved that transparent data is not underwritten risk. Both share a single negative space: there was no continuous layer watching validator configurations for stale permissions, oracle feeds for manipulation surface, leverage and tokenholder concentration, or bridge value against the security budget protecting it.

But the era's signature lesson runs deeper than nobody was watching. It is that crypto's answer to insecurity was to add controls (audits, multisig, formal verification, name-brand attestations), and each control, the moment it was trusted, became the attack surface. Nomad's auditable code was the looting manual. Multisig was Bybit's vector. The audit stamp became the thing that let capital stop asking questions. A control you trust without continuously re-testing is not a safeguard. It is a single point of failure wearing a safeguard's clothes.

None of this is an argument against verification. It is the case for a different kind. The error was never that crypto verified too much; it is that it verified once and treated the property as permanent. That distinction (verifiable versus verified, the data versus the discipline of continuously watching it) is the hinge the rest of this thesis turns on.

Two forces are now arriving simultaneously, and each one independently makes underwriting harder than crypto has ever managed.

The first force is tokenization. Real-world assets onchain more than doubled between the start of 2025 and mid-2026, passing $31.4B by May 2026 by rwa.xyz's count, a pace no prior DeFi subsector has matched. And that parenthetical, depending on whose methodology, is itself an underwriting datum: the leading trackers of this sector disagree with each other by a factor of two on how big it even is. BlackRock, Franklin Templeton, Ondo, and dozens of institutional issuers are racing to put bonds, equities, CLO tranches, and real estate onto public chains, and the buyers now showing up are sovereign wealth funds and the largest asset managers on the planet.

This is the best thing that has ever happened to crypto. From an underwriting perspective, it is also the most dangerous. Every tokenized real-world asset carries a promise: that a token onchain is backed by a real asset held somewhere offchain. That link depends entirely on trust frameworks between issuers, auditors, and custodians. The smart contract can run flawlessly while the offchain entity behind it quietly fails.

This is not a new risk. It is the central risk of all structured finance, and finance got it catastrophically wrong inside living memory. The 2008 crisis was, at its core, a failure to re-underwrite the assets sitting behind a security: AAA-rated mortgage paper whose underlying loans nobody re-examined. The token-versus-asset gap that every RWA depends on is the security-versus-collateral gap that detonated the global financial system.

The second force is AI that can find and exploit vulnerabilities at industrial speed. On April 7, 2026, Anthropic announced Claude Mythos Preview and declined to release it publicly, citing the danger of misuse. By Anthropic's own disclosures, the model autonomously discovered thousands of previously unknown vulnerabilities across major operating systems and browsers, including flaws that had survived decades of human review, and produced working exploits without human guidance. Rather than a general launch, Anthropic seeded it to a limited consortium through Project Glasswing, a defense-first program.

Mythos-class tooling compresses the attacker's timeline from years to hours and removes the headcount constraint entirely. The industry's own data already points the same way: the median time from vulnerability discovery to exploitation has collapsed from over two years in 2018 to hours today. A restricted preview buys time; it does not repeal the capability.

Two forces, converging. Offchain assets crypto might struggle to verify. AI-augmented attackers crypto cannot outrun. One makes the solvency side of underwriting exponentially harder; the other does the same to the security side. Both arrive at once, into a system that has historically failed at each in isolation.

The three eras describe when crypto failed. This act describes how: the recurring mechanisms underneath the history.

In any speculative bubble, leverage concentrates in three kinds of actors: those who can generate outsized returns, those who command significant liquidity, and the rails through which returns are generated, the protocols themselves. Each is a pressure point, and each fails in a characteristic way depending on the season. When demand surges, fraud risk spikes. When demand wanes, insolvency risk spikes as actors climb the risk curve to survive. And underneath every season, hacks remain a constant.

Fraud is the hardest risk to underwrite, because it means the information you were given was engineered to mislead you. Every other risk assumes good faith but incomplete data. Fraud assumes the data is a weapon.

This is why FTX humiliated the most sophisticated allocators alive. The diligence failure was not that Sequoia or Paradigm lacked access; it is that fraud defeats data-room diligence by construction: the hole does not appear in the documents the borrower controls. The Madoff parallel is precise: the SEC examined Madoff multiple times and missed it, while Harry Markopolos reconstructed the fraud from public return data and was ignored for nearly a decade. Wirecard ran the same way. EY signed clean opinions while the FT's Dan McCrum, working largely from public filings, was attacked for years before being vindicated. The pattern repeats in crypto exactly: the institutions with formal access miss it, and a lone analyst working from public data calls it early and is dismissed.

Crucially, fraud is not a CeFi-only phenomenon, and the curator collapses of 2025 prove it. Stream Finance presented as a transparent, composable DeFi protocol while depending on an opaque offchain fund manager, which post-collapse reporting alleges used protocol assets to cover its own trading losses. The onchain surface looked clean the entire time. Transparency at the contract layer is no defense when the lie lives one layer down, offchain, where the contract cannot see it. The digital-native version of Madoff's basement.

A hack does not stay a security event. It becomes a solvency event, and then a contagion event. Wormhole needed a $320 million backstop from Jump Trading or it would have been insolvent. DMM Bitcoin did not survive its hack at all. Every drained protocol is instantly an underwriting problem for everyone who lent to it, held its token, or accepted its receipts as collateral.

The Jump backstop is itself a TradFi maneuver, a private lender of last resort stepping in to prevent contagion, the same role the New York Fed played in organizing the $3.6 billion rescue of LTCM in 1998, and JPMorgan and ten other banks played in the $30 billion backstop of First Republic in 2023. The difference is that TradFi has institutions and a central bank explicitly mandated to perform that function; crypto relies on whether a profitable market maker happens to feel charitable that week. This is why security and solvency cannot be separated in practice: an exploit is just an insolvency that arrives in a single block instead of over a quarter.

In a downturn, macro liquidity leaves the system and yields compress. Ethical operators respond by shrinking or shutting down. Gamblers respond by reaching further up the risk curve to defend a return they can no longer earn honestly. This is the quiet failure mode, and 2025 was its definitive demonstration.

TradFi has a name for this exact behavior, reaching for yield, and treats it as one of the most reliable precursors to a blowup. It is what drove pension funds into structured credit before 2008, and it is what produced the UK's liability-driven investment crisis in September 2022: pension funds that had levered up to hit return targets faced cascading margin calls as gilt yields spiked, were forced into fire sales, and had to be rescued by emergency Bank of England intervention. That episode (leverage to manufacture yield, a sudden move, forced selling, contagion, a bailout) is the curator collapse of 2025 with different nouns.

The curator model concentrated this dynamic into a single role. Curators do not custody user funds, but they choose collateral and set risk parameters, and their fees reward yield. That incentive, in a low-yield environment, is an incentive to dance with risk. It was best described not by a crypto founder but by Citigroup's Chuck Prince in July 2007, on the eve of the crisis: as long as the music is playing, you've got to get up and dance — and we're still dancing.

On November 3–4, 2025, Stream Finance disclosed roughly $93M in losses tied to its offchain manager and halted withdrawals. Its yield-bearing token xUSD fell about 77% within a day, froze around $160 million in deposits, and propagated roughly $285M in interconnected debt across the ecosystem.

Three details make Stream the cleanest illustration of the entire thesis. First, the leverage was hidden: holders learned only after the collapse that xUSD reportedly carried something like $170 million in backing against roughly $530 million in borrowings, the same hidden-leverage surprise that turned LTCM and Archegos from private bets into systemic events. Second, the concentration was extreme: a small number of curators controlled the overwhelming majority of the relevant vault TVL, so a single failure was a systemic one. Third, and most damning, at least one prominent curator publicly acknowledged having identified the centralized counterparty risk during pre-listing due diligence, and listing the asset anyway because demand was too strong to pass up. Commit now or lose the round, ported intact from venture capital into onchain credit.