A vulnerability in Coldcard hardware wallets has been exploited to steal nearly $90 million in cryptocurrency, according to Galaxy Research. The flaw, which shipped in March 2021, reduced the entropy of seed generation to about 40 bits on some models, letting attackers guess private keys. The attack is still ongoing, and the tally of drained addresses keeps climbing.
The five-year-old bug
Coldcard wallets are designed to generate seeds using a hardware entropy source. But a preprocessor guard routed seed generation to a weak software PRNG called Yasmarang instead, collapsing the effective entropy to roughly 40 bits. That's a far cry from the 256 bits you'd expect. With only 40 bits, a determined attacker can brute-force the keys.
The flawed code shipped in March 2021 and stayed in publicly readable firmware for more than five years. Nobody noticed until someone started sweeping addresses. Attackers initially hit 500 addresses. Galaxy Research's count now stands at 4,585 addresses and nearly $90 million, with no sign of stopping.
Why the license change didn't help
Coinkite, the company behind Coldcard, had moved its firmware from a free-software license to source-available terms (MIT with a Commons Clause) after Foundation Devices used the code in a competing product. The idea was to keep others from copying it. But that move didn't increase protection at all.
The bug lived in code that a machine could read regardless of the license. The license change only changed the economics of finding the bug. It didn't remove the information. In the age of highly skilled AI, everything distributed is readable. Binaries can be decompiled, and obfuscation doesn't remove information. Trade secrecy in shipped software is just obscurity, and reverse engineering a product you lawfully possess is fair play.
AI found what auditors missed
Coinkite's working assumption is that someone used AI to comb the publicly available firmware and find the bug. That's plausible. An AI-assisted audit run weeks before the theft found nothing. But since the attack started, researchers have shown several frontier models locating the same flaw in minutes from a single prompt.
This isn't an isolated case. On July 20, mathematician Levent Alpöge posted a counterexample to Keller's Jacobian conjecture, open since 1939, using Anthropic's Claude Fable 5. The result was verified by Lean within hours. In May, an OpenAI model toppled the Erdős unit-distance conjecture, open since 1946. In late July, a 30-year-old graph-theory conjecture fell to four prompts. AI isn't just reading code; it's solving problems that have stumped humans for decades.
The new security reality
For hardware wallets and any shipped software, the lesson is blunt: if your code is distributed, assume it will be read. A license won't stop an AI. Obfuscation won't stop an AI. The only real protection is making the code actually secure, not hiding it.
Coinkite hasn't announced a fix yet. The attack is still running, and Galaxy Research's tally continues to climb. For Coldcard users, the question is whether their keys were generated after the vulnerable code was removed. As of now, there's no public statement on when a patch will arrive or how affected users should respond.
