Transcription
2026 has officially become the worst year for crypto security on record. And we aren't even halfway through it.
$450 million was drained from DeFi protocols across 145 separate instances in the first quarter alone. Another $66 million has evaporated in just the first 18 days of April. Total value locked has collapsed from $110 billion in January to roughly $82 billion as I record this video, a one-year low for the entire sector.
But the truly alarming part of the story is not the size of the losses, but the fact that all of this is unfolding at the exact moment that AI has become dangerously good at one specific thing, and that's finding bugs in code. Anthropic's latest model has autonomously discovered thousands of zero-day vulnerabilities, including one that has been hiding inside OpenBSD for 27 years. The cost to scan an entire smart contract for exploitable flaws using this same model, $1.22.
Today, we reveal why this represents the start of an entirely new era for software security and why crypto is just the first asset class in the crosshairs. My name is Lewis and you're watching the Coin Bureau.
Back in 2019, the meantime from vulnerability disclosure to active exploitation was 2.3 years according to data published by Immunifi. Now it's less than 24 hours. Median time to exploit brand new Common Vulnerabilities and Exposures, or CVEs, has fallen to under 5 days. 131 new CVEs are being disclosed globally every single day. And on the offensive side, adversary breakout time, the gap between an attacker gaining initial access and beginning lateral movement, has dropped to an average of 29 minutes, with some documented cases as fast as 27 seconds. Put simply, the operational tempo of attacks has accelerated by approximately two orders of magnitude in less than a decade.
And the two largest exploits of April 2026 illustrate exactly what this acceleration looks like in practice. On the 1st of April, Drift Protocol on Solana was drained for approximately $285 million in an attack that has been attributed by both Elliptic and TRM Labs to North Korea's Lazarus Group following a six-month social engineering campaign. Then on the 18th of April, the Kelp DAO Bridge was compromised for an additional $293 million, again attributed to Lazarus, this time by exploiting a single signature validator configuration on a Layer Zero adapter. Combined, these two incidences alone represent $577 million extracted by a single nation-state actor inside a single calendar month.
And here is why this is only the beginning of the story. On the 7th of April, Anthropic publicly announced what they are calling Claude Mythos preview, released under a restricted access defensive program known as Project Glass Wing. The model is not a general-purpose chatbot. It is a frontier system specifically designed for autonomous cybersecurity research, autonomous coding, and complex multi-step reasoning. And according to Anthropic's own disclosure, it represents what they describe as a step change in raw offensive capability.
The specific discoveries that have been confirmed are honestly extraordinary. One was a 27-year-old denial-of-service vulnerability in the OpenBSD TCP sack implementation, a bug that had survived nearly three decades of human review. A 17-year-old remote code execution flaw in the FreeBSD NFS subsystem was also found, and a 16-year-old vulnerability in FFmpeg, the multimedia framework that underpins a significant portion of the modern internet's video infrastructure. There are plenty of others, too. Anthropic claims that the model has identified thousands of additional high-severity zero-days across the broader software ecosystem. And as of mid-April, approximately 99% of those vulnerabilities reportedly remain unpatched.
In one specific controlled test against the Firefox 147 JavaScript engine, Mythos achieved a 72.4% exploit success rate, although independent analysis have correctly noted that this figure reflects a narrow test environment in which the model repeatedly weaponized two specific bugs. But the smart contract benchmark is where the story collides directly with crypto. Anthropic researchers working alongside the Matts Fellowship constructed something called Scone Bench, a test suite of 405 real-world smart contracts that had been exploited between 2020 and 2025 across Ethereum, Binance SmartChain, and Base. When frontier AI models were turned loose on these contracts, they produced fully working exploits for 207 of them, a success rate of 51.1%. The simulated stolen funds across that single benchmark exceeded $550 million. And on a separate test against 2,849 newly deployed contracts that had been considered secure as of October 2025, the same models discovered two completely novel zero-day vulnerabilities that no human auditor had previously identified.
And this brings us to the economic dimension that genuinely changes everything. According to Anthropic's own published cost data, the price of running an exhaustive AI-powered vulnerability scan on a single smart contract is approximately $1.22. Yeah, just $1.22. For comparison, a traditional human-led smart contract audit from a tier-1 firm costs anywhere between $50,000 and $500,000, and it takes weeks to complete. And inference costs are continuing to fall by roughly 23% every 2 months.
You might assume that the defensive side could simply match this asymmetry by deploying their own AI to patch vulnerabilities at the same speed. However, that assumption entirely ignores the fact that defenders need to be right every single time while attackers only need to be right once. Furthermore, Anthropic researchers have observed that the simulated revenue generated by AI-powered exploits has been doubling approximately every 1.3 months over the past year. Independent academic research published on arXiv has calculated that an attacker becomes profitable on AI-generated exploits at a target value of around $6,000, while defenders require detection budgets exceeding $60,000 to justify equivalent AI-powered audit infrastructure. That's a 10-to-one structural imbalance baked directly into the economics of the technology. And computational costs for generating these exploits dropped approximately 70% over a recent 6-month window, enabling roughly 3.4 times more attacks per unit of attacker budget. Put simply, the entire deployed surface area of every smart contract in DeFi history can now be scanned for less than the price of a single tier-1 audit. The defensive side cannot scale fast enough to keep up.
And this brings us to the half of the equation that almost nobody is talking about. Because while Mythos and its successors are getting brutally efficient at finding bugs, the same broader category of AI tools is also being used to write the code that contains those bugs in the first place. The 2025 Veracode Gen AI code security report analyzed over 100 large language models across 80 different coding tasks. The headline result was that approximately 45% of all AI-generated code introduced security vulnerabilities matching the OWASP Top 10, a widely used awareness list of the most critical security risks to web applications. In Java specifically, the failure rate exceeded 70%. A separate Stanford and NYU study tested 89 development scenarios and found that 40% of AI-generated code contained security vulnerabilities. The Veracode data also confirmed that AI-written code is 2.74 times more likely to contain cross-site scripting vulnerabilities than human-written code, with AI tools failing to prevent XSS in approximately 86% of test cases. Apparent research is even more alarming. By June 2025, AI-generated code was introducing more than 10,000 new security findings every single month, a 10-fold increase from December of 2024. AI code contains 2.5 times more high-severity CVEs, 322% more privilege escalation paths, and 153% more design flaws than equivalent code written by humans.
So, the compounding picture is genuinely alarming. Developers are using AI to ship code faster, but attackers are using AI to break that code faster. Both sides are accelerating, but only one side actually bears the cost when something fails. And the so-called "vibecoded" smart contracts, the protocols thrown together using AI assistance without rigorous formal review, have become the single easiest possible category for the target of AI-driven exploitation. The exact same model that wrote the contract can frequently identify the exact flaws that it introduced.
A live preview of where this is heading played out just 2 days ago. On the 19th of April, Versecel disclosed a sophisticated supply chain attack that began with a single context.ai employee's machine being compromised by Lumisaler malware. In February, the attacker used a captured OAuth token to pivot directly into Versecel's internal Google Workspace, then into the broader Versecel infrastructure. Versecel CEO Gulmo Ranch publicly described the attacker as, in his words, "significantly accelerated by AI," citing the velocity and deep architectural knowledge displayed throughout the breach. Crypto projects across multiple ecosystems immediately scrambled to rotate API keys and conduct emergency code inspections.
Now, it's tempting to see this as fundamentally a crypto security story. That assumption, however, completely ignores the reality of how modern software actually works. You see, OpenSSL is a cryptographic library that underpins the vast majority of the entire global internet, including web servers, mail gateways, industrial automation systems, and central bank settlement infrastructure and the SWIFT messaging network. In January of 2026, an AI system identified 12 separate zero-day vulnerabilities inside OpenSSL, some of which had been embedded in the codebase since the 1990s. CVE-2026-31790 is an RSA memory leak that affects banks and DeFi protocols simultaneously. CVE-2026-5194 is a digital signature forgery flaw with a CVSS score between 9.3 and 10, which remains active as I record this. The same shared cryptographic primitives that secure over $200 billion in DeFi total value locked also secure global payment networks, retail banking platforms, and central bank settlement layers. A vulnerability in one is a vulnerability in the other.
And this is precisely why between the 8th and the 10th of April, Treasury Secretary Scott Bessant and Federal Reserve Chair Jerome Powell convened an emergency closed-door meeting in Washington with the CEOs of America's largest banks, including Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo. The meeting was specifically focused on the systemic risk posed by Mythos-class AI capabilities to legacy payment infrastructure. The Federal Reserve Bank of New York has previously modeled cascade scenarios, concluding that a successful cyber incident on any of the top five US banks could destabilize the entire financial system. Bank of England Governor Andrew Bailey has publicly warned that this generation of AI has the unique capability to, in his words, "crack the whole cyber risk world open" by stringing together multi-step attacks that legacy patching cannot mitigate.
Now look, stories like this are moving faster than any single video can keep up with. New exploits, new model releases, new treasury meetings. Half of what I just covered didn't exist a month ago. So, if you want to stay ahead of this stuff in real time, not days later when it hits mainstream financial media, then come join us over at the Coin Bureau Telegram channel. It's where we drop breaking analysis the moment major exploits hit, share the on-chain data that we're tracking, and flag the protocols that we think are most exposed before the headlines catch up. It's completely free to join. The link is in the description below, and I would generally recommend it if you're holding any meaningful capital in DeFi right now.
Now, let's get back to the story. JP Morgan CEO Jamie Dimon has stated directly that AI will "almost surely make cybersecurity risk worse in the near term," even as the bank allocates $19.8 billion to its 2026 technology budget. And of course, we should absolutely trust the same banking sector that produced the 2008 financial crisis to gracefully navigate an offensive AI revolution that they openly admit they cannot keep pace with.
So the reason crypto attracts headlines while traditional finance does not is purely a function of disclosure architecture. Crypto losses are transparent on-chain. Every single dollar drained from a DeFi protocol is publicly visible, fully attributable, and immediately auditable by a researcher with a block explorer. Traditional finance operates under non-disclosure agreements, regulatory cover, and reputational incentives that strongly favor quiet remediation over public acknowledgment. The losses are not smaller, they are simply invisible.
And speaking of the institutional response, the FS-ISAC advisory issued just the other day on the 20th of April contains nine specific recommendations for hardening financial sector perimeters with the explicit warning that "traditional assumptions and approaches for vulnerability management are no longer valid." The FDIC issued reverse model risk management guidance on the 17th of April. The Pentagon has reportedly applied a formal supply chain risk designation to Anthropic itself.
This brings us to the only question that genuinely matters in the long run. Can the industry actually defend itself against this? Well, the honest answer is that the defensive side is real. It is well-funded, and it is absolutely structurally disadvantaged. CertiK launched its AI auditor in April with an 88.6% cumulative hit rate against 35 real-world Web3 incidents from 2026. Coinbase released Frosty the same month, an internal AI auditor with an F1 score 1.5 times higher than the second-best tool on the market, capable of completing a full audit in 1 to 2 hours at approximately 1/100th the cost of a manual review. Trail of Bits has transitioned to an AI-native model running 84 specialized agents that can identify up to 200 bugs per week on optimal engagements. Certora is embedding formal verification directly into the AI code generation loop. Hypernative monitors over 75 blockchains in real-time, detecting 300+ distinct threat types and partnering with TRM Labs to neutralize malicious transactions before they finalize.
But the asymmetry is structural, not temporary. The Rariable protocol was drained of $25 million in 17 minutes earlier this year despite having undergone 14 separate audits. Trail of Bits has publicly confirmed that human auditors were reviewing less than 20% of all deployed contracts by the end of 2025. And that figure can only get worse as deployment velocity continues to accelerate. Anthropic's own disclosure that 99% of the vulnerabilities Mythos has identified remain unpatched is essentially the entire problem in a single statistic. The discovery side has decoupled completely from the remediation side, and there is no clear path to closing the gap.
So if you are a user navigating this environment, the implications are reasonably specific. Cross-chain bridges remain the single highest risk category in DeFi, particularly any bridge running a one-of-one validator configuration. Liquid restaking tokens concentrate smart contract risk across multiple chains. And the Kelp DAO contagion has demonstrated exactly how that concentration propagates into lending markets. Newly deployed AI-assisted protocols with no audit history, no formal verification, and no real-time monitoring integration are the easiest possible targets in this new threat environment. The minimum baseline for any protocol holding meaningful capital should now include multiple recent audits from tier-1 firms, formal verification through tools like Certora or Helmos, real-time monitoring integration, time-locked upgrades of at least 48 to 72 hours, multiple signature thresholds above 1:1, an active seven-figure bug bounty program, and insurance coverage through providers like Nexus Mutual.
The AI versus crypto story is no longer a future risk to be priced in. It is a present-tense reality, and the data clearly shows that the attackers are decisively winning the early rounds. But the deeper truth here is that this is not really a crypto story at all. It is a software security story that is simply playing out first inside crypto because, well, of the public, transparent, immediate, quantifiable nature of on-chain losses. The same models breaking DeFi protocols today will eventually be turned on commercial banks, on payment networks, on central bank settlement systems, and on the critical infrastructure that underpins everything from electricity grids to hospital networks. So, the question is not whether crypto can defend itself against this new generation of offensive AI. The question is whether anybody can.
But what do you think? Will the defensive side of this arms race eventually catch up through formal verification, AI augmentation audits, and real-time monitoring? Or are we now living through the early innings of a shift that favors offense over defense for the foreseeable future? Let me know your thoughts in the comments down below. And if you want to understand exactly how the Kelp DAO bridge exploit unfolded and why composability turned a single signature compromise into a multi-billion dollar ecosystem rupture, well then you can check out our full deep dive on that incident right over here. Thank you all so much for watching, and I'll see you again very soon. This is Lewis signing off.