Crypto Hacks Are 'Down 47%' This Year. That's the Illusion. The Real Story in CertiK's Numbers: Attackers Stopped Hunting Bugs and Started Hunting Keys — at $13 Million a Wallet.

On July 17, 2026, Forbes published its read of CertiK's Hack3D report for the first half of the year, and the headline number looks like good news: $1,315,676,432 stolen across 344 on-chain incidents, down 46.8% from the same period a year earlier. CertiK CEO Ronghui Gu spends much of the interview taking that comfort apart. Nearly all of 2025's larger total came from a single event — the $1.45 billion Bybit theft in February of that year. Remove Bybit and the comparable 2025 figure is about $1.03 billion, which makes 2026 roughly 28% worse, not 47% better. The incident count tells the same story from another angle: 194 incidents in the second quarter, up from 145 a year before. The attacks did not slow down. They changed shape — fewer, in Forbes' phrase, but far more surgical.

Where they aimed is the part that should reorganize how you think about custody. The costliest category in the report is not smart-contract exploits, not rug pulls, not phishing. It is wallet compromise: more than $444 million in the half, averaging over $13 million per event — "by far the highest of any category," Gu wrote. Code vulnerabilities were the most common attack, 204 incidents, but they produced just $151.6 million — modest money, spread thin. The economics have inverted. Gu states the new equilibrium in one sentence: "Attackers are getting more return by going after key management, multisig governance, and operational infrastructure than by hunting for bugs in code." And his closing line is the whole report compressed: "A protocol can pass a flawless code audit and still lose millions because of a compromised admin key."

The two heists that defined the half prove the point, and readers of this blog have met them both. The Kelp DAO exploit took $291.3 million; the Drift Protocol breach took $285.3 million; together they account for nearly 44% of everything stolen in six months. Neither touched a line of audited code. As one widely shared analysis put it: Kelp's contracts were clean — a single compromised validator handling cross-chain message verification took everything down. Drift had no contract bug — a manipulated signing interface did the work. Gu's verdict: "Both were failures in operational and infrastructure security, not code vulnerabilities in the traditional sense." A decade of audits, bug bounties, and formal verification has genuinely worked — well enough that the rational attacker no longer reads your code. They study where your keys live, who can use them, and what has to be true for a signature to happen. Then they attack that.

The phishing numbers make the same argument at the level of individuals, and they are stranger than they first look. The number of phishing attacks fell by more than half, from 132 to 63 — and losses barely moved. Just four social-engineering operations produced $310 million, about 85% of all phishing losses, including a January attack in which one victim lost $284,785,689. One person. Read that number again and notice what it implies about structure rather than gullibility: it is only possible to lose $284 million in one phishing interaction if one deceived human, in one moment, holds the power to hand over everything. The lesson of the half-year is not "attackers are phishing less." It is that attackers have learned to stop casting wide nets and start aiming every resource at the small number of points where a complete secret and a single person coincide. Fewer, more surgical — the phrase describes the phishing column too.

There is even a preview of the next key holder to worry about, and it is not human. Gu flags AI agents with wallet access as "essentially a new kind of privileged key holder… except its decision-making can be manipulated through inputs in ways a human might catch and a poorly-guardrailed agent won't." It is already happening: the JaredFromSubway trading bot lost about $7.5 million in June after attackers deployed fake arbitrage pools that tricked its automated strategy into approving malicious contracts. Whether the key sits behind a person, a multisig, a validator, or a bot, the pattern of 2026 is identical — the secret is the target, and whatever stands between the attacker and the secret is the actual security model, no matter what the audit report says.

What does a report about nine-figure protocol heists mean for someone holding their own coins? More than it appears, because the logic scales down perfectly. Your seed phrase is your admin key. The audited code — your hardware wallet's secure element, the cryptography of BIP-39, the signature scheme — is the part that essentially never fails, exactly as protocol code held up in H1. What fails is the operational layer around the secret: where the complete phrase exists, who can reach it, what one deceived or coerced or careless moment can surrender. Every theme this blog has covered this year — the hidden camera, the wrench, the phishing letter, the drawer — is the retail edition of the same shift CertiK just quantified at the institutional level. The attacker's question is never "is the cryptography sound?" It is "where does the whole secret sit, and what is the cheapest path to it?" If the answer is one place, one person, one moment, you are the low-cost target in a market that has become ruthlessly efficient at finding them.

The defense the report's experts converge on is worth quoting because of how familiar it will sound. Ido Sofer, whose firm sells multi-party computation to institutions, frames security as raising the attacker's cost until the return on investment breaks: "If you put enough constraints and enough security rails, the cost is going to be higher than my neighbor… so probably they're going to go there because the ROI is lower." His mechanism: "Instead of having one key stored in one place, it's going to be distributed… harder to get to a place where you can actually move the funds." That is threshold thinking — the same principle this publication has argued from the beginning, and the same one institutions now pay for. This blog has been honest about where server-based MPC relocates trust for individuals: onto a vendor's infrastructure, terms, and continued existence. But the underlying principle — no single place, no single person, no single moment from which the complete secret can be taken — is simply correct, and H1 2026 is 344 incidents' worth of evidence. For a self-custodian, the vendor-free version is Shamir's Secret Sharing: encrypt the seed, split it into a threshold set of shares — say 3-of-5 — and distribute them across locations and people, so that below the threshold an attacker holds information-theoretically nothing, and no $13 million average payday sits behind any one door.

seQRets is the tool for exactly that layer, and the boundary is stated plainly: it cannot audit a protocol, secure a validator, or guardrail a trading bot — the institutional failures in CertiK's report have institutional fixes. What it does is take the one lesson those failures keep teaching, that the secret's surroundings are the attack surface, and apply it where you live: your encrypted seed phrase, split into QR-encoded threshold shares, distributed across the people and places you choose — no servers, no accounts, no vendor to trust or outlive, and no single point where the year's most profitable attack category can collect. The industry hardened the code, and the attackers moved to the keys. Make sure your keys are not standing where the code used to be.