tldr Coldcard devices are designed to use a true hardware random number generator (TRNG) built into the STM32 microcontroller chip to create highly unpredictable seeds with the full expected entropy (normally 128 bits for a standard 12-word BIP-39 seed). In March 2021, during a firmware update that involved migrating parts of the code to a new library (libNgU / MicroPython integration), the seed-generation path changed. Due to a configuration mistake with a build macro (MICROPY_HW_ENABLE_RNG set to zero) and an incorrect check in the supporting library (it only tested whether the macro was defined, not whether it was enabled), the code silently fell back to a weak software-based pseudorandom number generator (a Yasmarang-family PRNG from MicroPython) instead of the real hardware TRNG. That software PRNG was seeded mainly from non-secret, easily modeled device state such as the chip’s unique ID/serial number and internal timing/clock registers. As a result: • On affected Mk3 devices the effective entropy dropped dramatically (estimated around 40 bits under realistic attack assumptions). • Later models (Mk4, Mk5, Q) mixed in some additional entropy from secure elements, so they were somewhat better but still fell short of the target (roughly 72 bits of effective entropy). Cryptographic hashing cannot create more entropy than is present in the input, so the generated seeds became far more predictable than they should have been.