AINEWS
A five year old randomness flaw in Coldcard Bitcoin hardware wallets drained about 1,367 BTC (~$89M), and its CEO credits AI assisted review; the real shift is that AI now finds latent bugs faster than humans can patch.
Roughly 1,367 BTC, about $89 million at the time, was drained from Coldcard hardware-wallet users in three attack waves over the past week. Losses may climb to 2,055 BTC (~$130M) in a likely fourth wave that Galaxy Digital, a crypto-focused financial firm, flagged on Monday (RT, via Toronto Telegraph). The devices were never connected to the internet. The flaw lived in firmware Coinkite shipped in March 2021.
The bug is a seed-generation entropy flaw. Versions 4.0.1 through 4.1.9 on the MK2 and MK3 produced insufficient randomness for 12- and 24-word seed phrases when users did not add at least 50 independent private dice rolls or a strong unique BIP-39 passphrase, according to Coinkite's security advisory (Coinkite blog). That conditional matters. A user who followed the dice-rolls step had a far stronger seed than the firmware alone provided. The bug is a default-quality problem, not a guaranteed compromise of every device, and the conditional path is why some Coldcard users are unaffected.
The scope is broader than the headlines suggest. Coinkite's advisory and the coldcard.com downloads page cover Coldcard MK2, MK3, MK4, MK5, and Q devices, plus separate Edge tracks (Coldcard downloads). Patched firmware went out on July 31, 2026: MK2 and MK3 at version 4.2.0 or later; MK4 and MK5 standard at 5.6.0 or later; Q standard at 1.5.0Q or later; Edge tracks at 6.6.0X and 6.6.0QX. Standard and Edge are separate release tracks, and Coinkite explicitly warns against cross-applying them.
The "AI paradigm" framing is Coinkite CEO Rodolfo Novak's, not a forensic finding. Novak, in a public apology post, said AI-assisted code review surfaced the latent bug at a speed outpacing seasoned human auditors (Bitcoin Magazine). The available references do not contain independent proof that AI-assisted review actually discovered this specific bug; that is a vendor claim, not an established fact. The CEO's framing is marketing under pressure: the alternative is a five-year-old, conditional entropy flaw that human audits missed the first time.
Coldcard firmware is open source, and any latent flaw in that code is, in principle, equally readable to attackers and defenders using modern static analysis or LLM-assisted review. The theft is not a wholesale indictment of cold storage. "Cold" describes the network posture, not the code, and a wallet that never touches the internet can still lose funds if its seed-generation path is weak. The conditional dice-rolls mitigation also surfaces a recurring pattern in the wallet space: a feature sold as a power-user upgrade is also a safety net the default firmware relies on without enforcing.
The same logic applies beyond Coldcard. AI-assisted code review is now faster at finding latent bugs in long-lived, open-source security products than humans are at auditing or patching them. Any firmware, library, or open-source security tool quietly carrying a low-probability flaw for years is now a slow-motion liability. Attackers who run the same review first extract value. Defenders who run it after ship patches under pressure. The time gap between discovery and patch is the new attack surface.
Two things to watch: whether Galaxy's fourth-wave estimate of up to 2,055 BTC (~$130M) holds or revises, and whether anyone reproduces or rebuts the AI-attribution claim with a chain-of-evidence account. The dollar figure is easy to track. The AI-attribution claim, if it stands, applies to every other long-lived open-source security product.