A Coldcard randomness failure allegedly exposed thousands of Bitcoin addresses, triggering theft waves and a renewed debate over self-custody risk.
For five years, a single line of code sat inside Coldcard hardware wallets — devices Bitcoiners trusted to protect their life savings — silently undermining the very randomness their security depended on. Nobody noticed. Not the developers, not the auditors, not the AI models Coinkite paid to scan the codebase. Then, in a 41-minute window on July 30th, 2026, roughly $70 million in Bitcoin vanished from 1,195 addresses. The bill for that one misplaced function call had come due.
What followed was a cascade of revelations, four waves of suspected theft totaling nearly $130 million, and the biggest single week of Bitcoin ETF inflows since April — $850 million rushing toward institutional custody as self-custody's halo cracked.
The root cause is almost painfully mundane. In March 2021, Coinkite migrated some cryptographic code to use Bitcoin Core's libsecp256k1 library. During the migration, the seed generation pathway — the code that creates your wallet's master key — was rerouted from Coldcard's hardware random number generator to a software pseudo-random number generator called Yasmarang.
Why? Because of a compilation flag named MICROPY_HW_ENABLE_RNG. It was set to 0, but the code only checked whether it was defined, not whether it was true. The flag existed, so the compiler took the wrong path. No error messages. No crashes. Just wallets generating seeds that looked random but were not.
The practical result: seeds that should have had 128 bits of cryptographic entropy instead had roughly 40 bits on older Mk2/Mk3 devices, and about 72 bits on newer Mk4/Mk5/Q models. A 40-bit search space is something a motivated attacker can brute-force with commodity hardware. You do not need to steal the device. You do not need physical access. You just need to know the device's serial number and approximate boot time — both of which can be inferred or obtained.
Galaxy Research identified four distinct waves of suspected theft. The first struck on July 30th — 1,195 addresses drained in under an hour, transactions all using identical 30 sat/vB fees with no change outputs, the fingerprint of an automated script. The second and third waves followed similar patterns but with different consolidation addresses. By August 3rd, a fourth wave pushed the estimated total to roughly 1,815 BTC across approximately 5,294 addresses.
Notably, the first wave happened roughly 30 hours before Coinkite publicly disclosed the vulnerability. Someone had already found the bug and weaponized it.
The market's response was swift and structural. Bitcoin ETFs recorded over $850 million in net inflows in the week following the disclosure — the strongest week since April — even as Bitcoin's price barely budged, hovering around $63,000 to $64,000.
The irony is thick. Coldcard was built for the hardest-core Bitcoin self-sovereignty crowd — the people who lecture you about not your keys, not your coins. And now it is those same users flooding into BlackRock and Fidelity ETFs.
Coinkite disclosed that it had run AI models over its codebase weeks before the breach and the models found nothing. Yet after the vulnerability was made public, community members reported that Claude Code identified the RNG issue in about 8 minutes, and an offline GLM-5.2 model found it in roughly 20 minutes.
The lesson is uncomfortable: AI can miss catastrophic bugs when scanning passively, but it can also find them with astonishing speed when aimed properly. The tool is not the problem — the process is.
If you hold crypto, this story should rattle you — not because Bitcoin's cryptography is broken, but because security abstractions are only as good as their implementation. A hardware wallet being offline does not mean its seed was generated securely. Upgrading firmware does not fix seeds that were already created with weak randomness. And open source does not mean audited.
The practical takeaways are clear: if you used a Coldcard with affected firmware, generate a new seed on verified patched firmware and migrate your funds. Dice-roll entropy adds independent randomness that bypasses the device's RNG entirely. For large holdings, multi-signature setups using hardware from different manufacturers reduce single-vendor risk.