The Coldcard Hack rocked the foundations of the bitcoin community.
gettyOn July 30th, attackers began draining bitcoin from addresses linked to seeds generated by compromised “cold storage” wallets. The devices, called Coldcards, were manufactured by Coinkite, and were widely viewed as the best choice for those who are paranoid in the extreme about cybersecurity. The fact that it was Coldcard and not some other device that fell victim to this hack rocked the bitcoin world at its foundations.
Galaxy Research tracked at least 15 different attackers who have stolen about 2,000 bitcoins worth about $130 million from 4,385 addresses in just a few hours. The speed and scale of the attack amounted to the worst disaster in bitcoin’s history.
The obvious question after such a catastrophe is whether self-custody itself has failed. Self-custody is promoted by bitcoin enthusiasts because it protects against institutional corruption and counterparty risk. But with this great power comes great responsibility – namely, to be 100% sure that the hardware and software one uses for self-custody of bitcoin is not itself compromised.
Financial sovereignty remains worth pursuing, because custodial risk remains very real. However, the bitcoin community is now coming to terms with the fact that it takes much more technical skill on the part of individuals to be sure their bitcoin is safe in their own custody – an idea that cuts against the notion that anyone, no matter their age, location, or background, can be sovereign over their digital property.
What actually happened in the Coldcard bitcoin hack
Coldcard devices are designed to generate wallet seeds using a hardware chip called a “true random-number generator,” or TRNG. When used properly, this random number essentially hides your bitcoin “in plain sight” but within a space that is incomprehensibly large, and much to big for computers to search. To get a sense of the scale: guessing a bitcoin seed is like finding a single atom in a billion galaxies’ worth of matter.
However, the Coldcard firmware was bypassing the use of this chip, and instead using a software function that was much less random. A seed generated this way is as easy to find as a single atom inside a large virus – a task that modern computing clusters can execute within a few hours.
The flaw in the firmware was hard to find for humans, but easy for AI coding agents. With the recent maturation of powerful AI models for coding, malevolent actors pointed their compute at Coinkite’s open-source repository and found a vulnerability that had escaped human notice for 5 years.
According to Coinkite’s own account, a link-time error caused seed generation to resolve to MicroPython's deterministic Yasmarang pseudorandom-number generator instead of the intended hardware source. The firmware therefore bypassed a functioning hardware generator and produced seeds from a much smaller and more predictable set of possibilities.
The consequences varied by device generation. Coinkite says Mk2 and Mk3 seeds generated on firmware versions 4.0.1 through 4.1.9 may have had roughly 40 bits of effective entropy under current attack assumptions. A separate analysis by Block’s Bitcoin Engineering and Security team found that later Mk4, Mk5, and Q devices added some additional randomness from the secure chip, but the reseeding process retained only four bytes. Coinkite estimates those devices produced about 72 bits of entropy instead of the intended 128.
The reason those numbers matter is to provide assurance that the incident did not break bitcoin’s cryptography, seed phrase technology, or hardware wallets as a category. Attackers exploited a smaller set of possible keys created by one implementation. Once they could generate candidate seeds and compare the resulting addresses with public blockchain data, the attack became an offline search problem.
