A volunteer Bitcoin security team says AI-assisted code review turned up 4, 962 potential issues across 390 Bitcoin-related codebases in just 29.8 hours, a number that is eye-watering even by crypto’s usual standards.
- 4, 962 potential issues flagged across 390 Bitcoin-related projects
- 720 findings marked high or critical severity
- Only 21.4% of the findings have been reproduced so far
- The sweep followed the Coldcard firmware flaw and linked thefts of 1, 596 BTC, with a suspected total near 2, 055 BTC
Bitcoin Red Team says the review effort started after the Coldcard wallet exploit exposed a serious firmware bug and pushed Bitcoin developers to look harder at wallets, cryptographic libraries, and infrastructure software more broadly. The numbers are striking, but there’s a catch: AI can spot bugs fast, and it can also spit out a mountain of noise unless humans verify what actually matters.
That tension is the whole story here. On one side, AI-assisted review can help surface security problems at a pace manual auditing simply can’t match. On the other, a giant pile of “possible issues” is not the same thing as a giant pile of confirmed vulnerabilities. Security work still needs adults in the room.
What Bitcoin Red Team says it found
According to Bitcoin Red Team, the first 29.8 hours of AI-assisted review produced 4, 962 potential issues across 390 Bitcoin-related projects, with 720 classified as high or critical severity. The group also said 21.4% of the findings have already been reproduced, meaning they were independently confirmed rather than just flagged by software.
That reproduction rate is the part worth keeping front and center. AI is great at spotting suspicious code patterns, but it also sees ghosts. Without human verification, maintainers can get buried in false positives, panicked triage, and reports that sound dramatic but never survive testing.
Bitcoin developer Calle, who is part of the effort, said the team was “averaging on the order of 1 critical exploit per hour per person” and described the security situation as “extremely bad.” That may sound brutal, but it also reflects a simple reality: when one serious wallet flaw gets exposed, nearby code tends to get dragged into the spotlight fast.
The group said the campaign costs roughly $10, 000 per day to run. OpenSats is covering the expenses, while Kimi Moonshot has provided AI accounts and access to its Kimi K3 model. The team also asked the community to contribute computing accounts or digital tokens to keep the work going.
For anyone still pretending software security is mostly a vibes-based exercise: it isn’t. It’s expensive, tedious, and relentless. The cost of not doing it properly tends to be much higher.
The Coldcard flaw that triggered the sweep
The wider review effort began days after the Coldcard wallet attacks exposed a firmware flaw. Coinkite, the company behind Coldcard, later confirmed the root cause: a firmware change introduced in March 2021 caused affected devices to use a deterministic pseudo-random generator from MicroPython instead of the intended hardware random-number generator.
That sounds like technical gobbledygook, but the security problem is straightforward. A hardware random-number generator is supposed to source randomness from physical processes. A deterministic pseudo-random generator is algorithmic, which means it can be far weaker if it gets used in the wrong place. For wallet seed generation, weak randomness is not a minor defect. It is the foundation cracking.
Seed phrases are the recovery phrases that control access to a wallet. If a seed is generated with too little entropy, it may be easier to guess or reconstruct than users expected. And here’s the ugly part: updating firmware does not magically fix a seed that was already created with bad randomness.
Coinkite’s guidance is blunt on that point. The patched firmware protects future seed generation, but wallets created with vulnerable versions still need to be replaced with entirely new seed phrases on updated devices, with funds transferred to fresh addresses.
Coinkite says the issue affects Mk2, Mk3, Mk4, Mk5, and Coldcard Q devices in different ways. It also says TAPSIGNER, OPENDIME, and SATSCARD are not affected because they use different codebases. See the Firmware Update Advisory for Coldcard Mk2, Mk3, Mk4, Mk5 and Coinkite’s COLDCARD Security Disclosure History for the vendor’s own breakdown.
How weak was the randomness?
Coinkite says affected Mk2 and Mk3 devices generated roughly 40 bits of effective entropy in the vulnerable cases. For Mk4, Mk5, and Coldcard Q, the vulnerable firmware produced about 72 bits instead of the intended 128 bits.
For non-specialists: entropy is the randomness that makes a seed hard to predict. The more entropy, the safer the key generation. So when the number drops from 128 bits to 72 bits, or worse, to around 40 bits in some cases, that is not a cosmetic downgrade. That is the kind of thing attackers love.
Coinkite also says the risk is reduced if users created wallets with at least 50 fair, independent, private dice rolls. In plain English, that means the user supplied their own outside randomness during setup. If that was done correctly, the specific RNG bug is not considered a risk from this issue alone.
A strong, unique BIP-39 passphrase can also add another layer of protection. But it is not a magic shield. It adds a second secret, it does not repair a seed that was generated badly in the first place.
How much money has been stolen?
Galaxy Research says attackers stole 1, 596 BTC from approximately 7, 300 addresses across three confirmed attack waves tied to the Coldcard seed-generation flaw. It also identified 14 smaller incidents linked to the same weakness.
Galaxy also flagged a suspected fourth coordinated wave that could push total losses to about 2, 055 BTC if confirmed. Earlier blockchain analysis estimated that suspected fourth wave involved 448.7 BTC moving from 709 likely victim addresses.
That fourth-wave number should stay in the “suspected” bucket until further confirmation. There is a big difference between a verified theft count and an estimate that may still change.
Galaxy’s on-chain analysis also found that about 90% of the stolen bitcoin has remained unmoved. Another attacker was observed routing 64 BTC through a Bitcoin mixer, while the largest known attacker was holding 1, 159 BTC across seven addresses without moving the funds.
A Bitcoin mixer is a tool used to blur transaction trails by pooling and redistributing coins. Sometimes people use that for privacy. Sometimes criminals use it to make stolen funds harder to follow. Usually, the blockchain doesn’t care which excuse comes first.
For additional context on the scale of the losses, see Coldcard Hardware Wallet Flaw Linked to $70 Million, Coldcard firmware bug may have exposed Bitcoin holders to, Coldcard Firmware Bug Drained Bitcoin After Weak Seed, and Coldcard flaw tops $100M as David Schwartz warns.
Why this bug is such a nasty one
Hardware wallets are designed to keep private keys offline and reduce attack surface. That part of the security model still works, until the seed generation itself is compromised. If the seed is weak, everything downstream inherits the weakness.
That is why this incident is so serious. It is not some web app bug or a funny little UI glitch. It is a flaw in the birth process of the wallet. Once the seed is poisoned, the rest of the setup can be perfectly fine and still be built on bad math.
Coinkite says emergency firmware updates are available for all affected products. But users with wallets created on vulnerable firmware still need to generate new seeds and move funds. In other words: patch the device, then migrate the wallet. Do not assume an update alone makes old seeds safe.
Users who are unsure whether they used vulnerable firmware, whether they added enough independent dice entropy, or whether their passphrase setup is truly strong should treat the wallet as suspect and follow vendor guidance carefully. Self-custody is great until you start improvising around cryptography like it’s a weekend cooking show.
What the AI sweep really means
The Bitcoin Red Team figures are attention-grabbing, but they should not be confused with a fully verified vulnerability count. The group says 4, 962 issues were flagged, but only 21.4% have been reproduced so far. That means most of the findings still need proper human validation.
This is the dark side of AI-assisted security work: it can dramatically speed up discovery, but it can also flood teams with junk. A model that produces a dozen plausible bugs is useful. A model that produces a hundred plausible bugs, half of them nonsense, is just a new kind of admin nightmare.
Even so, there’s a real upside here. If one firmware mistake can expose seed-generation problems in a major hardware wallet line, it makes sense to sweep the rest of the Bitcoin stack for similar nonsense. Wallet software, cryptographic libraries, and infrastructure code all deserve hard scrutiny, especially when real money is on the line.
That’s the part people in crypto still sometimes forget: decentralization does not mean “no security problems.” It means the opposite. The burden shifts onto code quality, operational discipline, and users who actually move when a serious bug is disclosed.
For a broader look at the automated audit effort, see Bitcoin Red Team finds 4, 962 issues reviewing Bitcoin and Bitcoin Red Team finds 4962 issues reviewing.
Key questions and takeaways
-
What triggered the Bitcoin Red Team sweep?
The effort was launched after the Coldcard wallet attacks exposed a serious firmware flaw tied to seed generation. -
Are the 4, 962 findings all confirmed bugs?
No. Bitcoin Red Team says they are potential issues, and only 21.4% have been reproduced so far. The rest still need human verification. -
What should affected Coldcard users do?
Update to the latest firmware, then create a completely new seed phrase on the updated device and move funds to new addresses. -
Does a firmware update alone fix old wallets?
No. Coinkite says the update fixes future seed generation, but existing wallets made with vulnerable firmware still need to be migrated. -
How bad were the confirmed thefts?
Galaxy Research says attackers stole 1, 596 BTC across three confirmed waves, with a suspected fourth wave that could lift the total to about 2, 055 BTC if verified. -
Does using dice rolls or a passphrase make users safe?
Not automatically. Coinkite says at least 50 fair, independent, private dice rolls can add enough outside entropy to avoid this issue alone, and a strong BIP-39 passphrase helps, but those are mitigations, not excuses to ignore the warning.
The takeaway is simple and uncomfortable: Bitcoin security is only as strong as the randomness behind it and the firmware that generates it. When either one fails, the whole “trustless” story gets a lot more human, a lot more fragile, and a lot more expensive.
