AGENT SKILL

rng-entropy-audit

Cryptographic-hardware RNG + seed entropy audit. Scores captured RNG output for bias / repetition / low min-entropy (TRNG→PRNG fallback like the Coldcard 2021 regression), and scores seed material (BIP39 word count + passphrase bits) for attack cost. Triggers on 'RNG bug', 'TRNG disabled', 'PRNG fallback', 'entropy audit', 'hardware wallet', 'seed strength', 'Coldcard', 'BIP39 passphrase'.

License 许可
Apache-2.0
最近更新
2026-11-09
Stars
★ 0
Forks
0

01 前言 Front Matter

name
rng-entropy-audit
metadata
{ "subdomain": "crypto-hw", "when_to_use": "hardware wallet RNG audit, TRNG PRNG fallback detection, BIP39 seed entropy, seed passphrase strength, firmware regression against a prior sample, incident postmortem for a shipped device", "upstream_ref": "NIST SP 800-90B (entropy sources), NIST SP 800-22 (statistical tests), BIP39" }
description
Cryptographic-hardware RNG + seed entropy audit. Scores captured RNG output for bias / repetition / low min-entropy (TRNG→PRNG fallback like the Coldcard 2021 regression), and scores seed material (BIP39 word count + passphrase bits) for attack cost. Triggers on 'RNG bug', 'TRNG disabled', 'PRNG fallback', 'entropy audit', 'hardware wallet', 'seed strength', 'Coldcard', 'BIP39 passphrase'.

02 资源文件 2

03 SKILL.md 正文

RNG + Seed Entropy Audit Skill

You are auditing whether a shipped cryptographic hardware device (wallet, HSM, TPM, YubiKey-class token) actually delivers the entropy its data sheet promises. The two inputs you accept are:

  1. A captured RNG sample (hex or base64, at least 4 KiB recommended) pulled from the device's random-source endpoint under owner consent.
  2. A seed spec (BIP39 word count + optional passphrase-bit estimate) the owner supplies for a strength check on their existing wallet.

The two outputs are structured findings that either (a) certify the device / seed as meeting the 128-bit post-Grover strength target or (b) name the concrete remediation (add passphrase bits, rotate to audited firmware, migrate off the affected chip).

Threat model (fixed — do not renegotiate)

  • TRNG-to-PRNG silent fallback is the class of bug that took down Coldcard Mk2/Mk3 in the 2021 firmware regression. The RNG source quietly switches from a hardware entropy source to a deterministic software PRNG seeded from stable device state (serial, boot counter, flash pointer). Once one sample is available the state space shrinks to something a laptop can enumerate.
  • Under-truncated seeds. BIP39 12-word seeds provide 128 bits of raw entropy (128-bit security). 24-word seeds provide 256 bits raw but are typically rated at 256 bits of security only because they reference a checksum that doesn't multiply security. Additional BIP39 passphrase bits are additive: each character of a cryptographically-random passphrase adds ~5.9 bits (95-char printable ASCII), each dictionary word from a 4096-word list adds 12 bits.
  • Grover halves the effective bits of any brute-force target. A 128-bit seed is 64 pq-bits under Grover; below the NIST Cat-1 128-bit target. The remediation (BEFORE Q-day) is to add passphrase bits — every doubling of effective bits doubles the Grover search cost.

Operating principles

  1. Score before you speculate. Run rng_entropy_score.py --sample on any captured RNG output and rng_entropy_score.py --seed on any seed spec. Findings ship with structured severity — no free-form speculation.
  2. Never accept a sample from an unverified source. The value of the audit depends on the sample coming from the actual device under owner control. Ingest only samples with a chain-of-custody record in findings/crypto-hw/<utc>-capture/provenance.json.
  3. Never persist the sample or seed itself. Only the SHA-256 of the sample and the shape of the seed (word count, passphrase bits) reach the KG. The raw bytes and any recovered mnemonic stay in the operator's ephemeral workspace and are shredded post-audit.
  4. Recommend passphrase bits BEFORE recommending device replacement. Adding 8 – 16 bits of high-entropy passphrase is cheaper for the owner than migrating funds off a hardware wallet and buys time until a firmware fix or a PQC-native successor ships.

Severity rubric — RNG sample

Signal Severity Finding
min-entropy H_min ≥ 7.0 bits/byte AND monobit |dev| ≤ 4σ AND repeated-block ratio ≤ 0.01 INFO RNG output indistinguishable from ideal at this sample size.
H_min between 6.0 and 7.0 bits/byte OR monobit 4σ – 6σ LOW Detectable bias; increase sample size and re-audit.
H_min between 4.0 and 6.0 OR repeated 4-byte block ratio between 0.01 and 0.25 MEDIUM Structural bias present; likely non-cryptographic RNG. Do not use for key derivation.
H_min < 4.0 OR any all-zero / all-ones block ≥ 16 bytes OR repeated 4-byte block ratio > 0.25 HIGH RNG output is deterministic or has failed catastrophically. Do not use. Assume any key derived from this source is recoverable.

The repeated-4byte-block ratio is the primary Coldcard-class signal. A truly random sample of ~1024 four-byte blocks has essentially zero collisions in a 2^32 space (birthday bound), so any measurable collision rate signals a small effective state — the fingerprint of a software PRNG cycling in a tight loop.

Severity rubric — seed spec

Effective bits (post-passphrase) Severity Finding
≥ 128 pq-bits (≥ 256 classical) INFO Meets NIST Cat-1 post-Grover target.
96 – 127 pq-bits LOW Below Cat-1. Add 8–16 passphrase bits on next rotation.
64 – 95 pq-bits MEDIUM Vulnerable to a CRQC by 2035. Rotate within 2 years OR add passphrase bits now.
< 64 pq-bits HIGH Recoverable pre-CRQC by nation-state targeting. Rotate immediately AND add ≥ 32 passphrase bits.

Workflow

1. Load spec + sample from findings/crypto-hw/<utc>-capture/.
   Require provenance.json (device serial, firmware version, capture
   method, operator identity).
2. If a sample is present:
     python rng_entropy_score.py --sample <sample.hex> [--label <name>]
   Emit finding = { source, bytes, min_entropy, monobit_dev, longest_run,
                    repeated_block_ratio, severity, rec }.
3. If a seed spec is present:
     python rng_entropy_score.py --seed --words <n> --passphrase-bits <b>
   Emit finding = { words, raw_bits, passphrase_bits, effective_bits,
                    pq_bits, severity, migration }.
4. If both firmware versions are supplied (regression mode):
     python rng_entropy_score.py --sample <before.hex> --label before
     python rng_entropy_score.py --sample <after.hex>  --label after
   Compare the two verdicts; emit an additional finding of severity
   HIGH if the "after" verdict is worse than "before" (regression).
5. Persist findings to findings/crypto-hw/FIND-NNN.md via the shared
   finding-protocol skill; each MUST cite provenance.json + the exact
   firmware version.

Inputs

RNG sample (hex-encoded bytes):

$ python rng_entropy_score.py --sample sample.hex --label "coldcard-mk3-5.0.6"

Seed spec:

$ python rng_entropy_score.py --seed --words 24 --passphrase-bits 0

Regression pair (before / after firmware upgrade):

$ python rng_entropy_score.py --sample before.hex --label before
$ python rng_entropy_score.py --sample after.hex  --label after

What NOT to do

  • Do NOT run this skill on a captured sample without provenance.json. A sample from an unknown source proves nothing about the device; the audit collapses to "here is a byte string".
  • Do NOT publish or transmit the sample bytes. Only the SHA-256 hash, size, and derived metrics are safe to persist. The sample itself can be device-identifying via subtle biases.
  • Do NOT recommend a device replacement on a MEDIUM finding alone. Adding passphrase bits is faster, cheaper, and reversible.
  • Do NOT run this skill on a target the operator does not own or does not have written consent to audit. Cryptographic hardware audits on third-party devices are out of scope regardless of RoE authorization.

04 同仓库更多 Skills