Coldcard RNG flaw sparks debate over AI‑assisted security reviews
Coinkite CEO NVK defends its AI‑driven audit after a $116 million Coldcard exploit, while Bitcoin developer Luke Dashjr argues the vulnerabilities were discoverable by human reviewers.
Key takeaways
- Coinkite claims an AI model uncovered a random‑number‑generator defect that human audits missed.
- Bitcoin Core developer Luke Dashjr published nine concrete issues he had previously reported to Coinkite, disputing the AI‑only narrative.
- The vulnerability, traced to a 2021 firmware change, enabled the theft of roughly $140 million in Bitcoin and has prompted legal action.
Coldcard hardware wallets suffered a critical random‑number‑generator (RNG) flaw that allowed attackers to reconstruct private keys and steal an estimated $140 million in Bitcoin. Coinkite’s CEO Rodolfo “NVK” Novak has defended the company’s security process, saying a new AI model identified the defect after multiple human reviews failed to do so. Bitcoin developer Luke Dashjr publicly challenged that claim, releasing a nine‑point list of RNG concerns he had previously shared with Coinkite in private messages.
BTC / USD
AI‑driven audit versus human review
NVK posted on X that at least 15 external reviews—including paid audits of the Mk4 (2022) and Q (2024) models—had examined Coldcard’s hardware, firmware, and entropy‑generation code. He cited an OpenAI‑based “Cybersecurity Harness” that generated 85 candidate findings, arguing that only this sophisticated model uncovered the seed‑generation defect. According to crypto.news, NVK wrote, “The uncomfortable truth is that it took a new AI model to find what humans couldn’t.”
Dashjr’s nine‑point rebuttal
In a thread on X, Luke Dashjr listed nine specific issues he believes a competent reviewer could have spotted. The items range from a TODO comment in the RNG initialization routine to ignored seed‑error and clock‑error bits on the STM32 chip, and insufficient post‑processing required for cryptographic use. He also highlighted that the random_buffer check does not cover direct calls to rng_get, and that the firmware’s replacement RNG lacked runtime assertions. Dashjr emphasized that his findings were shared privately with Coinkite before going public and that he did not claim every item was exploitable.
Technical background of the flaw
The vulnerability originated from a March 2021 firmware change that moved seed generation from the hardware RNG to a software tool (Yasmarang) seeded by the chip’s serial number and a timer. This reduced effective entropy to as low as 40 bits on older Mk2/Mk3 devices, far below the 128‑bit security assumed by BIP‑39. Attackers were able to reconstruct private keys offline, leading to a series of thefts beginning July 30, 2026, in which roughly 1,816 BTC (about $38 million) were stolen from over 5,200 addresses, according to TRM Labs and Galaxy Research.
Legal and remediation response
Coldcard users have been instructed to generate new seeds, back up securely, and migrate funds using the latest firmware releases (5.6.3 for Mk4/Mk5 and 1.5.3Q for Q). Coinkite notes that firmware updates cannot repair previously generated weak seeds. A class‑action lawsuit has been filed in Ontario by wallet owner Amir Reza Naderi, alleging that Coinkite knew about the RNG issue since May 2025 and failed to warn customers. Coinkite denies the allegation and says a detailed technical post‑mortem is still in preparation.

Discussion 0