Bitcoin Red Team Critical Flaws: What 85 Bugs Mean

The Coldcard bug was the trigger. A flaw in how the popular hardware wallet generated randomness let attackers predict or recover private keys, and funds were actively stolen across multiple waves before a fix shipped. That kind of event — a widely trusted device failing at the one thing it exists to do — tends to make an open-source community ask an uncomfortable question: what else is sitting undiscovered in code that hundreds of thousands of people trust with their money?

Sixteen volunteers answered by building a roughly 170,000-line testing harness and pointing frontier AI models — including Kimi K3, GPT variants, Opus and GLM — at the Bitcoin ecosystem's codebase, spending about $40,000 in AI compute along the way. The pace was the point: 4,962 findings surfaced in the first 27.5 hours, growing to around 6,700 across 425 projects by hour 55. No manual review process moves that fast. That speed is also exactly why the 85 critical-flaw number needs a caveat most headlines are skipping.

AI code review is good at generating candidates and bad at knowing which candidates matter. Of the findings so far, only around 21% have been independently reproduced by a human — meaning roughly four in five "critical" flags haven't yet been confirmed as real, exploitable bugs rather than false positives or misreadings of intentional design. Eight have already been retired outright. Fewer than 5% of the 390-plus projects involved have even received formal disclosure yet, which is the standard first step before a bug becomes public or gets a fix. Calle, one of the team's co-leads, has publicly apologized to maintainers for the volume of unverified reports the sprint has generated, which is a telling signal about where the process currently sits: overwhelmed by its own output, not yet in control of it.

Severity also isn't spread evenly. Privacy and coinjoin tools showed the highest concentration of high-or-critical findings, around 24%, versus roughly 10% for cryptographic libraries and about 9.6% for hardware wallet firmware — the category that actually matters most for a self-custody user's day-to-day risk.

Here's the part that should actually change behavior: none of these 85 flaws are reported as currently being exploited. The one confirmed, active exploit remains Coldcard, and it predates this sprint — it's the reason the sprint exists, not one of its findings. The 85 are under private responsible disclosure, the normal process where researchers tell a project's maintainers before telling the world, giving them time to patch quietly. For most self-custody holders, particularly anyone using a mainstream, actively maintained wallet, the realistic near-term outcome is a wave of quiet patch releases over the coming weeks, not a new round of drained wallets.

That doesn't mean ignore it. It means the actionable move is boring but correct: keep firmware and wallet software updated as patches land, and treat any specific advisory naming your exact wallet model with real urgency rather than treating the 85-flaw headline as a reason to panic-move funds.

Maintainers of well-resourced, actively developed projects benefit — they get free, fast vulnerability scanning that would otherwise cost far more in time or contracted audits. So do users of those projects, eventually, once patches ship. The losers, at least in the short term, are smaller or under-maintained projects now facing a flood of unverified reports with limited volunteer time to triage them — exactly the maintainer fatigue Calle apologized for. If a genuinely critical bug gets buried under noise because a small team can't tell signal from false positive fast enough, that's the real near-term risk, not a mass exploitation event.

The team plans to open-source its testing harness, which would turn this from a one-off panic response into a recurring practice — arguably the most durable outcome of the whole episode. But the same tooling that let volunteers find bugs this fast is available to anyone, including attackers. Ledger's CTO has already made the obvious point publicly: open source and reviewed are not the same thing. The bull case here is that Bitcoin's security process visibly modernizes and the ecosystem-wide baseline rises. The bear case is that verification never catches up with disclosure volume, or that someone running the same AI tools independently reaches a live-exploitable bug in a widely used wallet before the volunteers or maintainers do. Watch the reproduction rate over the coming weeks, not the headline count — that's the number that will tell you whether this sprint made Bitcoin safer or just louder.