In this article
- TL;DR
- First, what a seed phrase is
- Entropy, in plain terms
- True randomness versus a formula pretending to be random
- The exact Coldcard entropy bug
- Who got swept, and who did not
- What Coinkite confirms, and what it does not
- The part that should worry every open-source project
- What to do right now
- Swiss holders, a note on tax
- The self-custody argument this reopened
- Where the CLARITY Act fits, and where it does not
- Bottom line
- Related reading
I work in the crypto self-custody space, and in my view this is the worst hardware-wallet failure yet: weak randomness in Coldcard firmware let attackers rebuild seeds and sweep Bitcoin.
TRM Labs counted roughly 594 BTC, close to USD 38 million at the time, leaving about 500 wallets in the first 25 minutes on 30 July 2026 (TRM Labs). That was only the first wave. By 14 August, Galaxy Research had confirmed with high confidence 1,778.84 BTC stolen from more than 8,600 addresses (Galaxy Research).
The victims were careful people. Galaxy Research describes the Coldcard as the device Bitcoin maximalists promoted as the most security-conscious hardware wallet (Galaxy Research). One victim told Galaxy he kept his Coldcard in a bank safe-deposit box, never connected it to the internet, and still lost 18.25 BTC in seven minutes (Galaxy Research).
TL;DR
A build error that shipped with Coldcard firmware in March 2021 made seed generation draw from MicroPython's software random number generator instead of the chip's hardware one (Coinkite technical deep dive). Seeds meant to carry 128 bits of randomness carried about 72 bits on Mk4, Mk5 and Q (Coinkite advisory), and Coinkite's preliminary estimate for Mk2 and Mk3 is about 40 bits (Coinkite technical deep dive). Attackers regenerated those seeds offline and stole funds. Seeds made with at least 50 private dice rolls were not at risk from this bug. A strong passphrase added a barrier, but Coinkite still tells passphrase users to migrate. If you generated a seed on affected firmware, migrate now.
First, what a seed phrase is
Skip this section if you already know how seed phrases work.
When you set up a Bitcoin wallet, the device generates a secret number, a very large random one. It then encodes that number as a list of ordinary English words, the "seed phrase," 12 words for 128 bits of entropy (BIP-39). Treat those words as the Bitcoin itself. Whoever holds them controls every coin the wallet has ever received, and there is no reset or support line to call.
The whole security model rests on one assumption: that the secret number is random, so nobody else can ever produce the same one. That assumption is what Coldcard broke.
Entropy, in plain terms
Entropy is a fancy word for "how much real randomness is in that secret number," measured in bits.
Each bit doubles the number of possible seeds. The table uses one RTX 4090 testing about 1.5 million BIP-39 candidates per second. That rate comes from the hashcat benchmark of 3,120.9 kH/s for PBKDF2-HMAC-SHA512 at 999 iterations, scaled to the 2,048 iterations BIP-39 uses.
| Entropy | Possible seeds | Time for one RTX 4090 to try them all |
|---|---|---|
| 128 bits | 2^128 (~3.4 × 10^38) | about 7 × 10^24 years |
| 72 bits | 2^72 (~4.7 × 10^21) | about 98 million years |
| 40 bits | 2^40 (~1.1 × 10^12) | about 8 days |
A correct 12-word seed carries 128 bits. NASA gives the Sun roughly 5 billion more years (NASA), so one card would need about 10^15 times the Sun's remaining life to finish. That is the number that makes self-custody safe.
Drop it to 40 bits and one card covers the whole space in about 8 days, and a thousand cards in about 12 minutes, at the same benchmark rate. That is the whole story: a seed you could not guess in the Sun's lifetime became one you can guess in a week.
The 72-bit row looks safe on one card. Block's engineers argue the real search can be far smaller, because for Mk4, Q and Mk5 only a 32-bit value from the secure elements reaches the generator, and an attacker who can pin down the timer state and call history faces at most 2^32 candidates (Block Engineering).
True randomness versus a formula pretending to be random
Hardware wallets are supposed to use a TRNG, a true random number generator: a hardware circuit built to produce unpredictable bits.
A PRNG, a pseudo-random number generator, is just a formula. Feed it a starting value and it produces numbers that look random but are fully determined by that starting value. Same input, same output, every time.
If the PRNG's starting value is something an attacker can learn or guess, the attacker can run the same formula and reproduce your "random" number exactly. Block found Coldcard's fallback was seeded from a fixed chip identifier, partly visible as the device's USB serial number, and from timer registers, none of them a secret (Block Engineering). The attacker did not have to guess 2^128 seeds. They had to guess a much smaller set of device and timing states, then regenerate the outputs.
The exact Coldcard entropy bug
Coinkite's technical deep dive gives the mechanism. In March 2021 Coldcard moved its elliptic-curve code to Bitcoin Core's libsecp256k1 through a library called libNgU. The board configuration set MICROPY_HW_ENABLE_RNG to zero, meant to disable MicroPython's software generator. The guard in libNgU checked whether the macro was defined, not whether it was nonzero, so the build linked MicroPython's software generator (Yasmarang) into seed generation. The hardware generator was in the binary, but seed generation never reached it.
Block's engineers, analyzing the firmware independently on 30 July 2026, date the vulnerable path to firmware 4.0.0, released on 17 March 2021 (Block Engineering). The flaw sat in the open-source code for more than five years.
The advisory also covers dice: "COLDCARD hashed the device-generated seed together with every dice roll entered through Add Dice Rolls" (Coinkite advisory). The device-generated half of that mix is the part that failed.
The damage by model:
- Mk2 and Mk3, firmware 4.0.1 to 4.1.9: the worst hit. Coinkite's preliminary estimate is an effective search space of about 40 bits (Coinkite technical deep dive).
- Mk4 and Mk5, seeds before standard 5.6.0 (or Edge 6.6.0X): about 72 bits instead of 128 (Coinkite advisory).
- Q, seeds before 1.5.0Q (or Edge 6.6.0QX): about 72 bits, per the same advisory.
A weak seed stays weak forever. Restore it onto a Trezor, a Ledger, a Sparrow hot wallet, anything, and it is still the same guessable number. Trezor said its own devices are not affected and told users who generated a wallet on a Coldcard and later restored it on a Trezor to create a new wallet and migrate (Trezor). The weakness lives in the seed, not the box you plug it into later.
Who got swept, and who did not
TRM Labs traced the first wave of roughly 594 BTC from about 500 wallets into a single consolidation address within 25 minutes, and saw three more waves in the following four days (TRM Labs). Galaxy Research later identified at least 33 further attacker footprints and concluded with high confidence that multiple attackers were active. It also found that not one theft transaction took funds from a multisig (Galaxy Research).
Two groups had protection:
Passphrase users, if the passphrase was strong. Coinkite's updated advisory says a strong, unique BIP-39 passphrase "adds an independent barrier," but "a short, common, patterned, quoted, or reused passphrase may be guessable" (Coinkite advisory). The passphrase, sometimes called the 25th word, is an extra secret you supply that the device never generates. It derives a completely different wallet, so an attacker who reconstructs your weak seed lands on the wallet without it. It survives the entropy failure because the broken component never produced it. A strong passphrase does not repair the seed, and Coinkite still says migrate as soon as practical.
Dice users. Coldcard lets you add your own randomness by rolling physical dice. Each six-sided roll contributes about 2.585 bits, so 50 rolls give at least 128 bits and 99 rolls give 256 (COLDCARD dice-roll math). Coinkite says users who entered 50 to 98 independent, private rolls got at least 128 bits from the dice alone, and it does not consider those seeds at risk from this bug (Coinkite advisory).
If you know your PIN but never set a passphrase or rolled dice, you are the exposed group.
What Coinkite confirms, and what it does not
Coinkite now states that a firmware bug weakened seed generation, that the devices were not hacked or remotely accessed, and that attackers exploited the weak seeds offline by regenerating the private keys and stealing funds (Coinkite advisory). Fixed firmware exists for every affected model and release track.
Its Security Status page calls itself the living postmortem for the incident and says it does not determine the cause of any individual reported loss. A separate detailed technical postmortem is still in preparation, as of the page verified on 1 October 2026 (COLDCARD Security Status).
The action you take does not depend on that postmortem. A seed from affected firmware without the dice exception is worth moving either way.
The part that should worry every open-source project
Coinkite wrote that because its source code has always been public, "we have to assume that someone used AI to review previous versions of our firmware and stumbled upon this issue." A few weeks earlier, it said, one of the best available AI models reviewed its code and found nothing serious (Coinkite technical deep dive). After the incident, Coinkite said it tested the code against further frontier models and none of them caught the bug (Coinkite).
Open-source firmware is a real strength, because anyone can read it and look for a backdoor. In my view it also means any attacker with a capable model and patience can replay years of commit history hunting for a subtle randomness mistake. Coinkite's own advice to other teams is to test AI review specifically against build and submodule boundaries, the place this bug lived (Coinkite).
What to do right now
If you have ever generated a seed on affected firmware, follow Coinkite's migration guide in order.
- Install the fixed firmware first. Mk2/Mk3 to 4.2.0 or later, Mk4/Mk5 to standard 5.6.0 or Edge 6.6.0X, Q to standard 1.5.0Q or Edge 6.6.0QX (Coinkite advisory). A fixed Mk2 or Mk3 can generate a safe replacement seed on its own, though Coinkite recommends moving long-term custody to a current Mk4, Mk5 or Q (COLDCARD Security Status). Updating never repairs the old seed. It only fixes new generation.
- Generate a brand-new seed and verify it. Use a second Coldcard on fixed firmware if you have one. Record and verify the backup and the wallet fingerprint before depositing, following the step-by-step wallet recovery guide if you need the exact restore flow.
- Send a test transaction, then move the rest. Verify the receive address on the device screen first. Repeat for every funded wallet on the old seed, including each passphrase wallet. Coinkite's guide warns against delaying an urgent migration to add a security layer you have not tested.
- If you want a dice-only seed on a Mk2 or Mk3, use at least 99 independent private rolls. The advisory describes this path as optional and advanced; it bypasses the device's generator and gives about 256 bits (Coinkite advisory).
- Never reuse the old seed anywhere. Its weakness is permanent. Restoring it on a different brand does not fix it.
- Ignore anyone who messages you offering help. No legitimate company will message you first or ask for your seed, passphrase, or dice rolls. The chapter on spotting scams hacks and common mistakes covers the playbook these accounts run.
For the mechanics of choosing a device and testing recovery before you deposit, see Wallets Staying Secure and the cold storage setup guide. For why brute force is normally hopeless, and why weak entropy is the exception that makes it trivial, see brute forcing BIP-39 seeds.
Swiss holders, a note on tax
Moving your coins from the old seed to a fresh one is not a sale. For private investors who are not classed as professional traders, gains on crypto held as private wealth are tax-free capital gains, and crypto holdings are declared for wealth tax at market value at the end of the tax period (ESTV working paper on crypto assets). The wallet that holds them does not change that. So there is no tax reason to hesitate on a migration, only the operational care of doing it right. If you are unsure how the move interacts with your cantonal filing, ask your cantonal tax office or a tax adviser before year-end. Background in declaring Bitcoin on your Swiss taxes.
The self-custody argument this reopened
The incident became ammunition within hours. Bitcoin commentator Guy Swann called it "the worst hit in bitcoin history to the most knowledgeable and 'properly secured' bitcoiners" (CoinDesk). That sting is real, because the victims followed the standard playbook.
The custodial camp pounced. Lorenzo Valente of ARK Invest said consumers "have traded counterparty risk for software risk, hardware risk, supply-chain risk, phishing risk, backup risk," and that "you are better off today holding funds across several publicly-traded exchanges or ETFs." David Lawrence of Amicus went further: "This is hugely damaging to the people who believe that 8 billion people will hold their Bitcoin in cold storage in the future. That dream is over. Done." Casa CEO Nick Neuman named the usability problem: "You just can't ask people to roll dice to be secure with your self custody. It's a non-starter for 99% of people." All three quotes appeared in the same CoinDesk report.
In my view they are half right. Pure single-signature self-custody does carry real, non-obvious risks, and this proved it. But a defense was in reach of the victims: dice at setup, or multisig where a single vendor's bug cannot drain you alone. Galaxy Research found that not one theft transaction took funds from a multisig (Galaxy Research). My answer to "one device failed" is not "hand your keys to an exchange," it is "stop relying on one device." That is Bitcoin self-custody done right, and the real tradeoff against custodial products is laid out in self-custody versus ETFs.
Where the CLARITY Act fits, and where it does not
The timing put a policy spotlight on the story. The CLARITY Act (H.R. 3633, the Digital Asset Market Clarity Act of 2025) is the US market-structure bill. The House text says a US individual "shall retain the right to" keep a hardware or software wallet for their own lawful custody of digital assets and to transact peer-to-peer (H.R. 3633, engrossed text). The House passed it 294 to 134 on 17 July 2025 (Clerk of the House, Roll Call 199). On 15 September 2026 the Senate rejected cloture on the motion to proceed, 49 to 50 (Congressional Record).
Coldcard is exactly the kind of hardware wallet that language is written to protect. But be precise about what the bill does not do. The House text contains no entropy standard and no firmware audit mandate; the words "entropy" and "firmware" do not appear in it (H.R. 3633, engrossed text). It protects your right to hold your own keys. It says nothing about whether the randomness inside your device is any good.
So the two intersect only in the debate they feed: how much of self-custody's risk should be left to open-source scrutiny and user diligence, and how much, if any, belongs in law. The hack did not stem from the bill, and the bill would not have prevented it. In my view, anyone pairing them as cause and effect is selling a narrative.
Bottom line
Bitcoin and BIP-39 held. A single vendor's firmware starved its own randomness for more than five years, and some of the holders who trusted it most paid for it.
The cryptography was never the weak point, and nobody brute forces 128 bits. The failure was a device handing out guessable numbers while displaying the same confident setup screen. The people who added independent randomness the device could not generate, at least 50 private dice rolls, kept their seeds out of the bug's reach.
If your seed came off an affected Coldcard, the move is not to panic and not to run to a custodian. Update, generate a fresh seed, move every funded wallet, and never reuse the old one. In my view self-custody survives this; trusting one device with one secret does not.
News and analysis, not financial, investment, legal, or tax advice. Figures, quotes, and firmware versions are drawn from Coinkite's advisory and status pages and the named research reports, as read on 11 October 2026. Loss totals were still being revised when Galaxy Research last reported. Verify your device's status against Coinkite's live Security Status page before acting.
Sources:
- Coinkite: Coldcard Security Advisory
- Coinkite: entropy issue technical backgrounder
- COLDCARD Security Status
- COLDCARD seed migration guide
- Block Engineering: Predictable RNG Fallback and 32-Bit Reseed in COLDCARD Firmware
- TRM Labs: Inside the USD 116 Million Coldcard Hack
- Galaxy Research: Coldcard Exploit Abates as Total Losses Climb
- Trezor: Coldcard vulnerability, Trezor devices are not affected
- H.R. 3633, Digital Asset Market Clarity Act of 2025 (govinfo)
Related reading
- Brute forcing Bitcoin BIP-39 seed phrases
- Bitcoin cold storage setup for 2026
- Step-by-step Bitcoin wallet recovery
- Bitcoin self-custody explained
New to Bitcoin? Start with Chapter 1.
Want the full picture? Read all chapters free or order the physical book.
