Dev-branch audit (docs/_archive/security-audit-dev-2026-07-22.md) re-found the
master issues (absent on dev) plus new ones in dev-only code. Applied here;
full-node build + test suite green; the subtle fixes were adversarially re-verified.
Ported from the master remediation (adapted to dev's code):
- bootstrap: reject zip-slip / path-traversal archive members (isSafeArchivePath)
before writing (S2-1). (dev already fail-closes on a missing checksum.)
- xmrig updater: fail closed when a signature is required but no key is pinned (F1-1).
- http_download.httpGetString: 16 MiB hard cap + MAXFILESIZE on the shared
metadata/price fetch (F1-2 / caps the updater + exchange paths at one site).
- rpc_client: explicit SSL_VERIFYPEER/VERIFYHOST (F4-1) and a 256 MiB response
cap in WriteCallback (F4-3).
- lite_connection_service: reject remote http:// lite servers, loopback only (L1-1).
Loopback is matched by a strict dotted-decimal 127.0.0.0/8 check (not a
startsWith("127.") prefix, which would wrongly accept 127.0.0.1.evil.com), with
userinfo/fragment stripping.
- lite controller: propagate encrypt/decrypt save() failure instead of reporting
success (F7-1).
- xmrig_manager: chmod(0600) the pool config before writing secrets (F5-2).
- app: clear the copied secret from the OS clipboard on shutdown (F3b-1).
- export_transactions: neutralize CSV/spreadsheet formula injection (F13-1).
- build pipeline: build-from-source lite backend + remove the self-attested
CMake signature gate (F15-1); pinned+verified appimagetool (F15-3/4);
verified Sapling params in setup.sh (F15-6); build.sh exits 0 on success.
(F14-1 empty-quoted-arg and F8-2 NUL-termination were already fixed on dev.)
Dev-only findings:
- rpc_client.callRaw: scrub the raw buffer + parsed tree (templated scrubJsonSecrets
for ordered_json) so console dumpprivkey/z_exportkey keys don't linger in freed
heap (N1-1).
- seed_wallet_creator: wipe the exported mnemonic on the failure path so a discarded
failed result never carries a live seed (W1-2).
- export_all_keys: write the plaintext key dump 0600 + atomically via
writeFileAtomically (U1-2).
- chat_database: restrict chat_messages.sqlite and its WAL/SHM sidecars to owner-only (C3-1).
Not done (need a decision, documented in the report):
- Chat header metadata (cid/z/p) rides outside the AEAD (C1/C2) — binding it is a
wire-protocol change requiring SilentDragonXLite interop review.
- Bootstrap lacks an offline-rooted signature (S2-2 residual) — needs signing infra.
- Console scrollback retains console-typed key-export output in plaintext (N1-1
residual) — inherent to an echoing console; would need output redaction.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
XmrigUpdater and DaemonUpdater carried byte-identical libcurl code —
httpGet, downloadToFile, and the write/progress callbacks — including the
security-critical TLS-verify (SSL_VERIFYPEER=1 / SSL_VERIFYHOST=2) and
max-body-size options, duplicated across two signature-verified
download-and-execute files that had to be hand-kept in sync.
Hoist that machinery into src/util/http_download.{h,cpp}:
- util::httpGetString(url, logTag)
- util::httpDownloadToFile(url, dest, logTag, maxBytes, onProgress)
threading progress + cancellation through a std::function (returning false
aborts the transfer) instead of the class `this` pointer. Each updater's
httpGet/downloadToFile become one-line forwarders that bind their own log
tag and archive-size cap (64 MiB miner / 256 MiB node, both preserved).
Every curl option is preserved byte-identically (verified by diffing the
setopt lines against the pre-change code — the only deltas are the
now-parameterized MAXFILESIZE cap and the XFERINFODATA payload). No behavior
change; the TLS/size-guard options now live in exactly one place.
Deliberately scoped DOWN per the audit's own guidance: the divergent
per-updater logic (release DTOs/parsers, checksum-table parsing, version
compare, and installResolved's per-member-SHA vs transitive-trust verify,
install strategy, and chmod scope) stays in each class — NOT merged behind a
base-class/virtual-hook layer, which would raise review cost on
signature-verified execute-code and risk a subtle security regression. The
shared-DTO/parser consolidation (~77 test refs of churn on stable code) is
left as a separate, lower-value follow-up.
Full-node + Lite build clean; ctest 1/1; source-hygiene clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
DanSandClaude Fable 5