Coldcard Flaw Tied to 1,816 BTC Moved From 5,200-Plus Addresses

About 1,816 bitcoin left more than 5,200 addresses between July 30 and Aug. 4 in transfers linked to a firmware flaw in Coldcard hardware wallets, according to on-chain records of the movements. The transfers carried a market value of roughly $116 million at the time — an average of nearly $22,000 per compromised address.

The episode is unusual less for its size than for its origin. Coinkite, the Toronto-based manufacturer of Coldcard devices, says an artificial-intelligence model identified the vulnerability that human reviewers had repeatedly failed to catch. CEO Rodolfo Novak said the codebase had undergone at least 15 reviews before the AI model flagged the flaw.

That claim invites scrutiny. Fifteen reviews missing a critical defect is either an indictment of manual auditing or a convenient framing, and Novak did not detail how the AI model was trained or deployed. What is not said matters here: the company has not explained when the flaw was introduced or how long affected firmware circulated before detection.

What exactly does the flaw do?

The vulnerability affects Coldcard firmware version 4.0.1 and later. It can bypass the hardware random-number generator, the component responsible for producing the entropy behind wallet seed phrases. A weakened or bypassed RNG undermines the assumption that a private key was generated unpredictably — the foundational security property of any hardware wallet.

The arithmetic is stark. At 5,200-plus addresses drained, the incident is not a single large theft but a broad sweep across many users, consistent with a systematic weakness rather than targeted phishing. Roughly 1,816 BTC is about 0.009% of bitcoin's total 21 million supply — small against the network, enormous against individual holders.

Did the developer community accept the account?

Not entirely. Some Bitcoin developers disputed the finding, and the dispute is not academic. James O'Beirne said he privately raised a concern about Coldcard's random-number generation as early as May 2025 — two months before the observed transfers began. If accurate, that timeline would complicate any narrative in which the flaw was unknown until an AI model surfaced it. Neither Coinkite nor O'Beirne has published the correspondence, so the sequence of who knew what, and when, remains unresolved.

The disagreement carries operational weight. Vendors rely on independent review to validate security claims; a manufacturer's characterization of its own discovery process, contested by a named developer, leaves users to weigh competing accounts without public evidence.

What can affected users actually do?

The remediation path is narrow. A firmware update can address the issue on new devices, but it cannot repair existing seed phrases — the entropy weakness, once baked into a key, travels with it. Users affected by the flaw must generate new seed phrases and migrate funds, a manual process with no vendor-side shortcut.

Put plainly: patched firmware protects the future, but any wallet created under the flawed RNG remains exposed until its owner acts. Coinkite has not, according to the available record, detailed a compensation mechanism for the roughly $116 million already moved.

What comes next?

The immediate burden falls on users: every holder who generated a seed on firmware 4.0.1 or later must treat that seed as suspect and rotate it, regardless of whether funds have moved yet. Watch how the O'Beirne timeline develops — public disclosure of the May 2025 exchange would either validate or dismantle the AI-discovery narrative. The broader market-structure question, whether hardware-wallet vendors will adopt AI-assisted auditing as standard practice after this episode, will play out over the coming release cycles.