Keylabs has completed an independent source-and-binary review of the remediation for the July 2026 COLDCARD seed-generation defect.

We are publishing the review before the full postmortem so the technical evidence is available now. This report does not change the migration guidance, repair an affected seed already generated on vulnerable firmware, or undo the losses suffered by customers.

Scope

Keylabs reviewed:

  • Mk5 5.6.2 and 5.6.3;
  • Q 1.5.2Q and 1.5.3Q;
  • the corresponding release-tag source;
  • the pinned libNgU and MicroPython submodules;
  • all four shipped, signed firmware binaries; and
  • Mk5 5.5.1 as a pre-fix negative control.

Keylabs checked each firmware image against Coinkite’s PGP-signed release hashes and independently disassembled the binaries.

The defect

The defect concerned entropy provenance.

COLDCARD’s firmware configuration set MICROPY_HW_ENABLE_RNG to zero. A guard in the pinned libNgU code used #ifndef, which tested whether that macro was defined rather than whether its value enabled the hardware RNG. The configuration was therefore accepted, and rng_get() resolved to MicroPython’s Yasmarang software generator instead of the STM32 hardware RNG.

The resulting output remained statistically uniform. Keylabs reproduced the defective generator and found that it passed the statistical health tests examined, including FIPS CRNGT, NIST SP 800-90B RCT/APT and SP 800-22.

Those tests examine output characteristics. They do not establish where the entropy originated. Detecting this defect required verification of the resolved path in the compiled firmware.

Keylabs did not independently establish Coinkite’s earlier estimates of approximately 40 or 72 bits of effective entropy.

Controls reviewed

Keylabs verified four controls against the faulty linkage:

  1. A value-checked compile-time guard rejects an invalid RNG configuration.
  2. The MicroPython software RNG object is removed from the build and the Yasmarang symbol is poisoned.
  3. The build-time rng-code-check verifies that the linked relocation chain terminates at the board’s hardware rng_get().
  4. A boot-time self-test reaches the hardware path before the VM starts and stops the device if that check fails.

The reviewed firmware also:

  • passes the full 32-byte SE1/SE2 digest to the boot reseed rather than truncating it to 32 bits;
  • uses SHA-256 Hash_DRBG seeded from 1,024 hardware-TRNG bits;
  • XORs a fresh hardware-TRNG word into each DRBG output word; and
  • conditions each seed with sha256d(TRNG || SE1 || SE2), using fresh TRNG and SE1 input together with SE2’s static per-device value.

Binary verification

The disassembly confirms that the reviewed firmware reads the STM32 hardware RNG data register.

The pre-fix Mk5 5.5.1 binary instead contains the Yasmarang software generator and does not reach the hardware peripheral through that path. Keylabs also found that the RNG-path C code is byte-identical within the 5.6.2/5.6.3 and 1.5.2Q/1.5.3Q release pairs.

Within the reviewed scope, Keylabs found the remediation sound, confirmed it in the shipped binaries, identified no remaining item required to close this vulnerability class, and concluded that the July 2026 defect is fully closed.

This was not a complete audit of every COLDCARD feature, model, component or release track. Edge firmware was outside the review.

The complete report is available as a PDF.

Report SHA-256:

c97b71ad3dfe6d06fc7d1295838799f691e58c5a021e3fb304de2215589e7f2c

If a seed was generated on affected firmware, installing an update is not sufficient. Create a new seed using corrected firmware and move the funds according to the current Security Status and migration guidance.

Our full incident postmortem will address the chronology, impact, contributing controls and remaining work.