Coldcard CEO says AI found the bug humans missed, but developers say they warned him first

Coinkite CEO NVK said it took a new AI model to find the seed-generation flaw behind the Coldcard exploit, which was linked to about $116 million in Bitcoin thefts since July. Bitcoin developers, including Luke Dashjr and James O’Beirne, countered that they had raised concerns about the code before the thefts.

Coldcard exploit puts earlier security reviews under scrutiny

NVK said at least 15 documented external reviews or security-research efforts had examined Coldcard hardware, firmware, tools related to random-number generation or release controls before the defect emerged.

He identified two paid third-party product reviews: one covering the Mk4 in 2022 and another examining the Q in 2024.

NVK also cited Rob Ham’s OpenAI Cybersecurity Harness review, which he said produced 85 candidate findings before the incident. Several findings led to fixes, according to his account, but none identified the seed-generation defect.

“The uncomfortable truth is that it took a new AI model to find what humans couldn’t,” NVK wrote on X.

That assertion became a central point of disagreement over the Coldcard exploit. Critics questioned whether the absence of a finding in those reviews justified the broader claim that human researchers could not identify problems with the random-number generator.

The scope of the reviews also matters. Work covering hardware, supporting tools or release procedures does not necessarily establish that every part of the seed-generation process underwent a comprehensive audit.

Developers challenge AI claim and handling of warnings

Bitcoin developer Luke Dashjr disputed NVK’s characterisation, saying he had privately communicated findings to Coinkite and was now making them public.

Dashjr qualified his disclosure, however. He said his examination had been brief, rather than a full audit, and did not claim that every concern represented a confirmed vulnerability.

“I’m not claiming these can all be weaponised, or that they’re definitely all bugs,” Dashjr wrote.

Those qualifications leave an important distinction in the Coldcard exploit debate: identifying a concern in code is not automatically the same as demonstrating an exploitable failure.

Other participants focused on the company’s approach to quality assurance. Alex Waltz accused Coinkite and NVK of discouraging scrutiny, while another participant, Kenneth, argued that periodic reviews of an entire codebase differ from quality checks integrated throughout development.

Shinobi defended NVK, saying reviewers had examined the entropy system without discovering the defect. He challenged critics to demonstrate a reproducible failure and argued that privately expressing concern did not amount to a security disclosure.

The competing accounts have made the Coldcard exploit discussion as much about disclosure standards and review practices as about the underlying technical flaw.

Earlier entropy concerns resurface after Coldcard exploit

Posts by Bitcoin developer James O’Beirne added another dimension to the exchange.

Writing on August 4, O’Beirne said he began examining Coldcard firmware in May 2025 to determine where its random-number generator obtained entropy. He questioned the use of the libngu library, raised concerns about hardcoded constants in its software fallback and recommended linking directly against libsecp256k1.

According to O’Beirne, he was told the configuration was correct for physical devices and that any problem would already have become apparent. He acknowledged that he did not pursue the matter rigorously.

In an August 5 follow-up, he described receiving another account of a warning dating to 2021, separate from an earlier Telegram discussion.

“It’s possible that 3 independent people warned them about this and they neglected it every time,” O’Beirne wrote.

His statement remains an allegation about those communications. The supplied reporting does not independently establish that three people disclosed the eventual Coldcard exploit or that each warning demonstrated the same underlying defect.

Theft estimate dates to August as postmortem remains pending

TRM Labs’ August 5 analysis described approximately 1,816 BTC stolen from more than 5,200 addresses across four waves beginning July 30, citing Galaxy Research’s running tally.

The roughly $116 million estimate therefore relates to the earlier thefts, rather than a newly reported loss arising from November’s argument.

According to that analysis, the Coldcard exploit originated in a March 2021 build configuration error that selected a weak software random-number generator. On older devices, effective seed strength reportedly fell to as little as 40 bits instead of the intended 128, allowing attackers to reconstruct keys without physical access.

Subsequent reporting tracked some stolen funds moving through THORChain into Ether. An August investigation update also said information had been passed to U.S. authorities, but it did not establish a confirmed attacker identification or recovery.

Coinkite describes its published security record as a living postmortem while stating that a separate detailed technical report remains in preparation. Its page distinguishes targeted independent checks from a complete audit of every corrected firmware binary.

The advisory also states that firmware updates cannot repair existing affected seeds. Its migration guidance calls for generating a new seed, verifying the backup and receiving address, making a small test transfer and then moving the remaining funds.

That limitation remains central to the Coldcard exploit response: correcting software does not restore the security of keys generated from an already weakened seed.