Luke Dashjr says Coldcard’s RNG flaws were manual finds, contradicting Coinkite
Bitcoin developer Luke Dashjr has contradicted Coinkite’s claim that only AI could detect flaws in Coldcard’s random-number generator, posting a nine-point list of vulnerabilities he says he flagged privately months earlier. For institutional custodians and self-custody users, the disclosure exposes gaps between Coinkite’s AI-assisted security review and what manual code inspection can find, raising questions about the wallet’s fitness for high-value holdings.
- Luke Dashjr identified nine RNG code issues in Coldcard firmware, including a TODO note, missing error checks, and weak post-processing on the STM32 chip.
- Coinkite released firmware 5.6.1 on August 20, 2026, adding seed-error checks, discarding 12 hardware words after recoverable errors, and switching to SHA-256 Hash_DRBG.
- Estimated losses from the July 2026 seed-generation breach exceed $140 million across roughly 500 wallets, with a class action lawsuit pending in Ontario Superior Court.
- $140M+ Cumulative losses from Coldcard seed-theft incident during July 2026
- 500 Wallets compromised in 25-minute theft window on July 30, 2026
- 40 bits Real randomness in Mk2/Mk3 seeds versus 128 bits BIP-39 assumes
Dashjr’s thread, first reported by Cryptopolitan, directly challenges Coinkite co-founder Rodolfo Novak’s framing of the vulnerability as something only artificial intelligence could uncover. Dashjr said he had raised the concerns with Coinkite privately but has no evidence they were fixed before the July 2026 breach that saw attackers steal approximately 594 bitcoin, then worth $38 million, in a single 25-minute window.
Dashjr flags TODO note and missing STM32 error checks in seed initialization
The first item on Dashjr’s list targets the rng_init function that initializes the Mk4’s RNG. He noted a TODO placeholder in the code and said that some STM32 chips require the first sample to be discarded, a step Coinkite’s driver skipped.
The driver also ignores the chip’s seed-error and clock-error status bits, Dashjr wrote, despite the STM32 manual calling for additional post-processing when the RNG is used for cryptographic purposes.
Dashjr’s concern about deterministic output also cuts deeper: if the STM32 became deterministic, he asked, why did the secure element’s randomness not collapse below its full 256-bit entropy? The question suggests a gap between how Coldcard assumed its layered entropy sources would reinforce each other and how they actually performed under fault conditions. Coinkite’s security update later confirmed that firmware now checks seed-error state, discards the next 12 hardware words after a recoverable error, and verifies at boot that rng_get() reaches the hardware read path, measures that directly address Dashjr’s flagged issues.
Firmware patch introduced novel RNG that Dashjr says required full retesting
Dashjr flagged the random_buffer function, which detects repeated RNG output, as inadequately tested when compared with direct calls to rng_get.
He also criticized Coinkite’s replacement code as essentially a novel RNG that needed complete cryptographic validation. The old PRNG was disabled only at build time, not at runtime, leaving Dashjr concerned that an RNG timeout in the wrong context could trigger undefined behavior, possibly a crash that would brick slow devices.
Coinkite’s patch moved seed generation away from the software Yasmarang tool (which relied on the chip’s serial number and a timer) back toward hardware entropy, but Dashjr’s point about validating novel designs under production constraints remains unresolved.
Breach origin and migration path leave institutional custody decisions unfinished
The theft began July 30, 2026, when Coinkite disclosed that a firmware change made in March 2021 had moved seed generation off the hardware RNG to the Yasmarang software generator. Mk2 and Mk3 seeds contained only around 40 bits of real randomness instead of the 128 bits BIP-39 assumes, allowing attackers to reconstruct keys offline.
No evidence has emerged that the Coldcard devices themselves were hacked or remotely accessed; instead, the attacker used the weak seeds to regenerate private keys and move funds from compromised addresses.
Coinkite’s updated firmware addresses the immediate technical failure but does not retrofit safety to existing vulnerable seeds. Users must create new seeds on the fixed firmware and migrate funds, a process Coinkite’s support team has been managing at scale. Plaintiff Amir Reza Naderi’s pending class action in Ontario Superior Court alleges Coinkite knew about the seed-generation problem from May 2025 and withheld disclosure from users, a claim Coinkite denies. The legal outcome will define whether the company’s internal knowledge of a vulnerability prior to public disclosure constitutes negligence toward custody holders.
The CCS read. Dashjr’s thread proves that human code review and AI-assisted review are not interchangeable, his nine points caught genuine weaknesses that Coinkite’s own AI tools missed weeks earlier. Institutional custodians should read this as a baseline: any wallet claiming AI-driven security without parallel manual audits by named external reviewers is overstating its assurance model. The question is whether Coinkite’s new firmware fixes are complete, which only sustained independent testing will answer.
On his X post, Dashjr did not say whether Coinkite has since fixed each of his nine items or whether he intends to file a separate security disclosure. Naderi’s class action is scheduled for hearing at Ontario Superior Court; the outcome will signal whether courts hold hardware wallet makers liable for pre-disclosure knowledge of design flaws. Institutional custodians awaiting Coinkite’s response to Dashjr’s specific list should also track whether Coinkite publishes a point-by-point rebuttal or confirms adoption of each fix.