Jonathan Goodman’s mnemonic words were written down and stored in safes and safety-deposit boxes. The device that generated them had never touched the internet. The cable was never plugged in, the QR code never scanned by anything connected to the wider world. Every visible precaution was observed with great care. Then, on the evening of July 29, 2026, the Canadian investor watched 18.25 Bitcoin leave his wallet over seven minutes. The sum was worth roughly 1.6 million Canadian dollars. He thought he had done everything right, but the money was now gone.
Goodman, like others, had put his trust in Coldcard, a hardware wallet for cryptocurrency. The natural instinct, when confronted with a theft like the Coldcard hack that hit him and many others, is to look for the intrusion: the broken window, the intercepted message, the insider. The Coldcard incident offered no such satisfaction. The secret was never stolen in transit; it was simply never secret enough.
What remains is not the absence of trust but its redistribution.
A hardware wallet is a small device that generates and guards the cryptographic keys controlling a Bitcoin balance. The device does not hold the coins. The coins exist as entries on a distributed ledger maintained by thousands of computers. What the device holds is the authority to move them, encoded as a string of words derived from a number that is supposed to be random. The entire architecture of private ownership rests on the proposition that the number your device chose when you pressed the button could not have been predicted by anyone else on earth. However, the Coldcard’s number could be.
A tale of self-deception
In 2021, Coinkite, the Toronto company that manufactures the Coldcard, migrated its cryptographic library. The migration was routine, and the new library was reputable. However, during integration, a function call that was supposed to reach the device’s hardware random-number generator instead resolved to a software fallback, a deterministic pseudorandom-number generator built into MicroPython. A hardware generator harvests physical noise, the kind of irreducible unpredictability that makes a secret genuinely singular. A software fallback produces numbers that look disordered but are fully determined by their inputs. Feed it the same inputs and it will produce the same outputs, every time, on every machine. The function returned bytes of the correct length in the expected format. The mnemonic phrase looked random, but it was not.
On older Coldcard models, the effective search space collapsed to roughly 40 bits, meaning an attacker had to consider about a trillion candidates rather than the intended hundreds of trillions. On newer devices, secure-element entropy raised the figure, but independent analysis found that only 32 bits of genuinely secret material survived in a critical reseeding operation. A trillion sounds like a large number, but it is not when the adversary has rented a cluster of GPUs and the public blockchain serves as a free answer key against which every guess can be checked in milliseconds.
The thefts were automated. On July 30, roughly 1,083 Bitcoin moved from 1,196 addresses in 41 minutes. Subsequent waves targeted thousands more. By early August, estimates ranged from 1,596 to over 2,000 Bitcoin drained across thousands of addresses, a sum approaching $130 million. The attackers did not need to know who owned the wallets or where the devices were stored. They needed only to regenerate the small universe of secrets those devices could ever have produced, derive the corresponding addresses, compare them against the blockchain, and sign. The protocol saw valid signatures, which is the only authority it recognizes.
RELATED: A shocking storage exploit bankrupts Bitcoiners — with lessons for us all
Velishchuk/Getty Images
The Coldcard had occupied a particular niche in Bitcoin’s self-image. It was the serious person’s wallet, the one recommended on the forums and podcasts that prized operational security, air-gapped signing, open-source firmware, and a philosophy of radical self-reliance. Its marketing vocabulary was a compressed manifesto: sovereignty, verification, trust minimization. Two weeks before the mass thefts, Coinkite published an essay arguing that offline status alone was insufficient and that real security required architectural separation, trusted displays, secure elements, and reproducible builds. The essay ironically enumerated the locks on the doors in a house whose foundation was already fissured.
Hidden is not secure
Bitcoin was designed to eliminate the need to trust institutions. You do not need to trust a bank to honor its ledger because the ledger is maintained by mathematics and distributed consensus. However, the self-custodying individual must trust or personally verify the firmware on the device, the compiler that built the firmware, the library the compiler linked, the symbol the linker resolved, the entropy source the symbol reached, the chip that generated the noise, the build system that packaged the result, the reviewers who examined the code, and the update channel that delivered the fix. One is trusting an entire institutional apparatus compressed into a plastic rectangle the size of a playing card.
“Do not trust; verify” is the community’s canonical imperative, but verification requires a division of labor. The user verifies an address on a screen. An auditor inspects the firmware. A reproducible-build operator compares binaries. A laboratory assesses an entropy source. No single person does all of these things, and almost no one does more than the first. What remains is not the absence of trust but its redistribution, away from a chartered institution with a complaints department and toward a dispersed, informal, largely volunteer apparatus of maintainers, reviewers, and forum participants whose oversight, in this case, missed a five-year-old defect in the most consequential function the device performs.
Jonathan Goodman’s safe was not breached. His backup was not photographed. His mnemonic words were not intercepted. The words themselves were simply a member of a knowable set, a predictable output dressed in the costume of randomness. The distinction between concealment and unpredictability turned out to be critical. While the Coldcard was offline, the institution that made the key never was.
Read the full article here
