Coldcard Bug Linked to Possible $71.1M Bitcoin Sweep; Coinkite Floats AI Theory

A long-running software bug in Coldcard hardware wallets may have enabled an attacker to reconstruct recovery seeds and drain 1,128.4717 BTC, roughly $71.1 million at the time of measurement. Coldcard maker Coinkite said it suspects artificial intelligence may have helped uncover the issue, though it has not presented proof of attribution. The incident drew attention after about 594 BTC (around $38 million then) moved on July 30 from roughly 500 single-signature Bitcoin addresses in a tight window of about 25 minutes. Follow-on blockchain analysis suggested a broader sweep: researchers estimated 1,082 to 1,196 addresses could have been hit over roughly 41 minutes. A public tracker, the Coldcard Sweep Watch dashboard, later pegged the total at 1,128.4717 BTC when bitcoin traded near $63,044. A screenshot taken at 8:30 a.m. ET on Aug. 1, 2026 showed that figure, which rose an hour later to 1,128.6633 BTC. Most of the stolen funds were consolidated into a single address holding hundreds of bitcoin, with much of the balance remaining largely unmoved. Analysts said the affected addresses shared a key trait: their recovery seeds were generated on Coldcard devices produced by Canada-based Coinkite. Recovery seeds are word lists that control wallet access; anyone who can reconstruct the seed can typically move funds without the physical device. Coinkite issued an urgent advisory warning that certain seeds generated on Coldcard devices could be weak due to a flaw in how randomness was produced. The highest-risk group includes Mk3 units running firmware version 4.0.1 (released around March 2021) and later. Coinkite said the issue may also affect seeds created on some Mk4, Mk5 and Q devices before emergency firmware fixes were released. The company said Tapsigner, Opendime and Satscard are not impacted because they use different software. The core problem centered on entropy quality. Secure seed generation depends on strong randomness so seeds cannot be guessed. For the most severely affected Mk3 devices, researchers estimated effective randomness may have dropped to about 40 bits instead of the intended 128 bits. That reduction dramatically shrinks the search space, making it feasible for a well-resourced attacker to test candidate seeds offline and match derived addresses against the public Bitcoin blockchain. Coinkite said some newer devices may have achieved roughly 72 bits of effective randomness due to additional unpredictable hardware input, improving resistance but still falling short of the intended strength. Coinkite said the vulnerability traces back to a build-time configuration mistake that survived for more than five years. Two functions with similar output formats were available: one used a hardware true random number generator, the other relied on a weaker software method inherited from MicroPython. Coinkite intended to disable the MicroPython path, but a software check only verified that a configuration label existed, not that its value was set to zero. That allowed the weaker function to be selected without triggering compilation errors. The company emphasized that updating firmware today does not fix a seed that was generated under the flawed code. Users who may have created seeds during the affected period must generate an entirely new seed on corrected firmware (or another secure device) and move funds to new addresses derived from that seed. Some users may have been protected if they added substantial external randomness when creating their seed, such as at least 50 independent dice rolls. A strong BIP39 passphrase could also have raised the difficulty of reconstruction. Multi-signature setups requiring approvals from multiple independent devices may have prevented a single compromised seed from authorizing transfers. Coinkite CEO Rodolfo Novak issued a public apology, saying the company takes full responsibility and is working on fixed software, technical write-ups and user support. Coinkite also advanced the idea that AI may have been used to spot the bug in older publicly available firmware. Novak wrote that AI-assisted code review can surface latent flaws faster than even highly experienced experts, warning developers that if firmware is open-source or has ever been public, it should be assumed to be scrutinized by both attackers and defenders. Modern AI coding tools can ingest large repositories and flag suspicious interactions among configuration settings, functions and security assumptions, including weak random number generation and cryptographic key handling. Some observers reported using top AI models to locate or explain the Coldcard weakness after the underlying randomness issue was identified, underscoring how accessible AI-assisted analysis has become. That said, it does not demonstrate the original attacker used AI. Coinkite also acknowledged that its own AI-based review did not catch the flaw before the theft, highlighting that AI tools are not guaranteed to surface every critical defect and can depend heavily on prompts, code context and human interpretation. Security specialists critical of the AI framing argue the root cause was a conventional engineering failure: a configuration error of a known type that standard code review, testing or audits focused on seed generation could have detected earlier. Others see the positions as compatible: human error created the opening, while AI could have reduced the cost of finding or exploiting it. For Coldcard users, Coinkite urged verifying seed creation timing and following official guidance to install corrected firmware, generate a fresh seed and migrate funds, while remaining alert for phishing and fake support messages. Longer-term questions include the final amount taken, whether the attacker can be identified, and whether AI played a decisive role. The episode is also expected to increase pressure on hardware wallet makers to strengthen entropy testing, audit build configurations and continuously re-examine legacy code with both human experts and adversarial AI tools.