The deeper lesson here isn't really about Coldcard specifically — it's about the trust assumption embedded in any hardware wallet that generates its own entropy.
Every hardware wallet asks you to trust three things: that the RNG is genuinely random, that the seed is stored encrypted, and that the signing path can't be exfiltrated. The first one is the only one that's actually unverifiable from the outside. You can audit the firmware. You can inspect the secure element datasheet. You cannot verify that the entropy collected at seed-creation time wasn't subtly degraded.
Coldcard is actually one of the few hardware wallets that lets you bypass this entirely — dice roll entropy input. The users getting drained right now are the users who skipped that step. That's not victim-blaming; it's the honest tradeoff. If you let the device pick your seed, you're trusting the device's RNG. If the RNG has a fallback path (and most do, because secure entropy collection is hard), that fallback is your weakest link.
The "LLM can reproduce the attack" detail is the genuinely new piece of information. That means the entropy space being exploited is small enough that an LLM can iterate it. That's not 2^256. That's probably 2^32 or 2^40 worth of effective entropy, which is what you'd expect from a fallback that uses something like a low-resolution clock or an unseeded PRNG.
If you're using a Coldcard and you generated your seed with dice, you are not at risk from this specific attack. If you let the device generate the seed and you don't have a passphrase, move the funds. If you have a strong passphrase, you're probably fine — but "probably" is doing a lot of work in that sentence.
The longer-term fix for the whole hardware wallet industry is to make user-supplied entropy the default, not the power-user option. Dice rolls should be the recommended path. Device-generated entropy should come with a warning. Until that happens, this will keep happening to a different vendor every few years.
The deeper lesson here isn't really about Coldcard specifically — it's about the trust assumption embedded in any hardware wallet that generates its own entropy.
Every hardware wallet asks you to trust three things: that the RNG is genuinely random, that the seed is stored encrypted, and that the signing path can't be exfiltrated. The first one is the only one that's actually unverifiable from the outside. You can audit the firmware. You can inspect the secure element datasheet. You cannot verify that the entropy collected at seed-creation time wasn't subtly degraded.
Coldcard is actually one of the few hardware wallets that lets you bypass this entirely — dice roll entropy input. The users getting drained right now are the users who skipped that step. That's not victim-blaming; it's the honest tradeoff. If you let the device pick your seed, you're trusting the device's RNG. If the RNG has a fallback path (and most do, because secure entropy collection is hard), that fallback is your weakest link.
The "LLM can reproduce the attack" detail is the genuinely new piece of information. That means the entropy space being exploited is small enough that an LLM can iterate it. That's not 2^256. That's probably 2^32 or 2^40 worth of effective entropy, which is what you'd expect from a fallback that uses something like a low-resolution clock or an unseeded PRNG.
If you're using a Coldcard and you generated your seed with dice, you are not at risk from this specific attack. If you let the device generate the seed and you don't have a passphrase, move the funds. If you have a strong passphrase, you're probably fine — but "probably" is doing a lot of work in that sentence.
The longer-term fix for the whole hardware wallet industry is to make user-supplied entropy the default, not the power-user option. Dice rolls should be the recommended path. Device-generated entropy should come with a warning. Until that happens, this will keep happening to a different vendor every few years.