Skip to content

Security: CreativeMayhemLtd/provcheck

Security

SECURITY.md

Security policy

Reporting a vulnerability

Email security@creativemayhem.com with the subject line provcheck security. Please do not open a public issue or pull request for a suspected vulnerability — give us a chance to ship a fix before it becomes a target.

Expect an acknowledgement within 72 hours. For severe issues we will coordinate a disclosure timeline with the reporter; for low-severity issues we typically fold the fix into the next regular release and credit the reporter in the commit body (opt-out on request).

PGP / Signal / encrypted reporting on request.

Supported versions

The latest published vX.Y.0 tag on GitHub Releases is the only supported version. We don't maintain backport branches. Earlier tags receive no security updates — if you're pinned to one, upgrade to track the current release line.

Iteration tags (vX.Y.Z where Z > 0) land on the dev repo as commit anchors during the pre-release coverage push and do not have published GitHub Releases; they are not "supported" in any form and are not intended for installation.

The release line is fast-moving by design (the project pre-dates v1.0); a security fix landing on main becomes the next patch tag, usually within 24 hours of the fix landing.

Tolerated transitive advisories

Several cargo audit / Dependabot advisories appear against provcheck that we cannot fix from this repository. They sit deep in transitive dependency chains where the immediate upstream parent has not yet migrated, or where no fixed version of the affected crate exists. Rather than ignore them silently, this section documents each one with our reasoning so contributors don't re-litigate them on every audit pass.

Re-evaluate the moment any of these get an upstream fix. The release-prep checklist runs cargo audit in both the workspace and app/src-tauri; if any of the IDs below disappear from that output, bump the affected dep and remove the corresponding row here in the same commit.

Workspace (Cargo.lock)

ID Crate Severity Path Why tolerated
RUSTSEC-2023-0071 rsa 0.9.x medium (5.9) c2pa → rsa Marvin Attack timing sidechannel. No fixed version exists upstream. Reachable only when verifying RSA-signed C2PA manifests; ECDSA / Ed25519 chains (the modal case, including everything provcheck-kit produces) are not affected. The rsa crate has acknowledged the issue but no fixed version has shipped — track the advisory page for status changes.
RUSTSEC-2026-0387 crypto-bigint 0.5.5 low c2pa → ecdsa → elliptic-curve → crypto-bigint Timing-variability in operations intended to be constant-time. Fixed in crypto-bigint 0.7.6, but reached only through elliptic-curve 0.13.8 and ecdsa 0.16.9, the ECDSA stack that c2pa 0.90.22 pins. crypto-bigint 0.7 requires elliptic-curve 0.14, which c2pa has not adopted, so there is no in-range bump. Same reachability class as the rsa row above: exercised only while verifying signatures on a local host, not against an attacker-controlled timing oracle. Re-evaluate when c2pa's ecdsa / elliptic-curve stack moves to crypto-bigint 0.7.
RUSTSEC-2024-0370 proc-macro-error unmaintained c2pa → static-iref → static-regular-grammar Build-time proc-macro, no runtime surface. Pulled by c2pa's iref / static-iref IRI-handling stack; the v1.5.0 tract 0.23 bump removed the older tract-stack path that also carried it. No fixed release of proc-macro-error exists; the row drops when static-regular-grammar migrates to proc-macro-error2 or c2pa's iref chain updates.
RUSTSEC-2026-0173 proc-macro-error2 unmaintained age → i18n-embed-fl → provcheck-sign Build-time. The age crate uses i18n-embed-fl for localised error messages; we use age for at-rest encryption only (passphrase-protected ES256 keys), so the i18n surface is not load-bearing for us. Wait for age to either drop i18n-embed-fl or for i18n-embed-fl to migrate off proc-macro-error2.
RUSTSEC-2026-0374 basic-toml 0.1.10 unmaintained age → i18n-embed-fl → i18n-config → basic-toml Build-time only. Parses i18n.toml inside the i18n-embed-fl proc-macro that age uses for localized error strings (the same chain as the proc-macro-error2 row). No runtime surface. Drops when i18n-config migrates off basic-toml or age drops i18n-embed-fl.
RUSTSEC-2026-0253 lru 0.16.4 unsound atrium-common → atproto attestation stack Unsoundness advisory (fixed in lru 0.18.2), pinned by atrium-common 0.1.4, which is already at its latest release. Unsound-classified, not a vulnerability with a demonstrated exploit path, and the lru use sits inside atrium's identity-resolution caching, fed by our own resolver flow rather than attacker-shaped keys. Bump lands automatically when atrium moves; re-check each audit pass.
RUSTSEC-2026-0371 / -0370 / -0372 zbus_names / zvariant / zvariant_derive unmaintained Linux keyring (secret-service D-Bus) Unmaintained advisories for the zbus D-Bus stack. Present in Cargo.lock but compiled only on Linux (the secret-service keyring backend); not in the Windows or macOS build tree (cargo tree -i zbus_names finds no path here). No fixed release. Drops when the keyring / zbus chain updates.
RUSTSEC-2026-0367 / -0368 / -0369 zip 8.6.0 medium c2pa → zip Three filesystem-extraction advisories in ZipArchive::extract / ZipStreamReader::extract: a crafted archive can place symlinks whose targets escape the destination directory and have Unix permissions, including SUID/SGID bits, applied through them. Fixed in zip 9.0.0 but unreachable here: c2pa 0.90.22 pins zip ^8.6.0 (verified: cargo update -p zip --precise 9.0.0 fails to select a version), so no c2pa release on zip 9 exists to bump to. zip is pulled only by c2pa and is never called directly by provcheck; the vulnerable path is extraction-to-filesystem, which the verify and sign flows do not perform on caller-supplied input, and the symlink, SUID, and SGID vectors are Unix-only and moot on the Windows release target. Re-evaluate when c2pa moves to zip 9.

Dependabot-only advisories (no RustSec ID, so not surfaced by cargo audit; tracked via GitHub Dependabot):

  • glib (GHSA-wrw7-89jp-8q8g, medium) — unsoundness in the Iterator / DoubleEndedIterator impls for VariantStrIter. Fixed in 0.20.0. Transitive via the same Tauri gtk-rs 0.18 stack described below; a major bump pinned by that generation, not runtime-reachable for us. Moves when the gtk stack moves.

Tauri app (app/src-tauri/Cargo.lock)

The desktop GUI inherits everything above, plus a long tail of unmaintained advisories from tauri 2.x's transitive use of GTK3 on Linux. The whole tail has a single root cause: Tauri 2.x's Linux backend (wry → webkit2gtk) is built on gtk-rs 0.18, and the gtk-rs maintainers marked the GTK3 bindings unmaintained when GTK4 became mainstream. Tauri's published migration path is wry's gtk4 backend; we move when Tauri does.

The unmaintained advisories in this group:

Plus a handful of unrelated unmaintained-only advisories carried by deeper transitive deps:

All of the above are tracked-and-waiting-on-Tauri-2.x's-gtk4-move. When the Tauri release that ships the gtk4 backend lands, bump the Tauri dep, re-run cargo audit, and remove rows from this section to match what actually cleared.

Update process

When cargo audit reports a NEW advisory ID that is not in the table above:

  1. Decide if a fix is reachable from this repo (direct dep bump, transitive bump via cargo update -p <crate>, or feature toggle).
  2. If yes, fix it in the next patch release.
  3. If no — document it here in the same commit that's adding it, with the same shape as the existing rows (ID, crate, severity, path, why tolerated, what fix path it waits on).

When an existing tolerated advisory clears (the rust-sec database or upstream ships a fix, or the dep tree shifts to drop the affected crate entirely):

  1. Bump the relevant dep in Cargo.toml / app/src-tauri/Cargo.toml or run the appropriate cargo update -p <crate> to refresh the lockfile.
  2. Re-run cargo audit in both Cargo.lock and app/src-tauri/Cargo.lock and confirm the ID is gone.
  3. Remove the corresponding row from this file.
  4. Ship the change as a patch release with the cleared ID called out in the commit body.

Scope

provcheck is a signing + verification tool. The most security- relevant surfaces are:

  • provcheck-sign — private key custody (OS keychain or age- encrypted file). The age-file backend is the only place we write long-lived secret material to disk, and the keychain backend delegates custody entirely to the OS. Both backends use authenticated encryption / OS-level primitives; we never roll our own crypto.
  • provcheck-publish — atproto session credentials in session.json (file mode 0600 where the OS supports it). The refresh JWT is the long-lived secret; access JWTs rotate every ~15 minutes.
  • provcheck-platform — fetches app.provcheck.signingKey records over HTTPS to atproto PDSes. Cache lives under the platform- appropriate data dir; never executes downloaded content.
  • The watermark detectors run ONNX inference via ort (onnxruntime 1.20 with the load-dynamic feature; the platform-specific CPU shared lib is bundled in the release archive). v0.7 phase 7b-followup migrated from tract-onnx to ort for the image-modality decoder; the audio detectors (silentcipher, AudioSeal, WavMark) followed in the same line.
  • Detector weights ship as download-on-demand DLC from the weights-v1 GitHub Release on the public mirror, NOT as include_bytes! blobs. This is the v0.8 phase 8a change that dropped the kit binary from ~143 MB to ~22 MB. The download + install flow:
    1. The kit binary embeds a compile-time MANIFEST table (crates/provcheck-weights/src/manifest.rs) of every weight: family name, variant name, file name, public-mirror URL, SHA256 hash, and expected size in bytes.
    2. On first detector call, provcheck-weights::load_if_cached looks for the weight in the platform cache dir (dirs::cache_dir() / provcheck). If missing, the verifier surfaces Error::NotCached and refuses to run; the operator installs via provcheck-kit weights install <family> which downloads to a tempfile, hashes it, compares to the MANIFEST SHA, and renames into the cache atomically only if the hash matches.
    3. On subsequent calls, the cached file is re-hashed against the MANIFEST entry before being handed to ort. A post-install bit-flip on disk surfaces as SHA mismatch and a clean error, not a silent substitution.
  • Model-substitution attack surface: an attacker would need to either (a) compromise the binary's compile-time MANIFEST table (covered by the platform's binary-integrity story — GitHub-signed release artefacts on Windows / macOS, distro package signatures on Linux), or (b) produce a weight file that collides with the SHA-256 hash of a legitimate weight (computationally infeasible). The DLC delivery model does NOT introduce a new substitution surface beyond what the binary's integrity already covers.
  • Windows binary integrity: Windows artefacts in a v*.*.0 release are Authenticode-signed by an SSL.com OV code-signing certificate held by Creative Mayhem UG (haftungsbeschränkt). Signed binaries carry an RFC-3161 timestamp so signatures remain valid past the cert's expiry. Verification via Get-AuthenticodeSignature on the target machine; the signing chain terminates at SSL.com's Code Signing Intermediate CA RSA R1. Iteration tags (vX.Y.Z, Z>0) skip the release matrix and therefore skip signing — do not treat an iteration- tag download as a supported install. Full signing procedure in docs/release-process.md § code-signing.

Things we explicitly do NOT do:

  • Auto-update on the user's behalf.
  • Phone home with usage telemetry.
  • Send file content over the network. Verification is local; attestation queries send the creator's DID or handle to their PDS to fetch the signing-key record, nothing more.

If you find a way around any of these promises, that is a security issue and we want to hear about it via the reporting channel above.

There aren't any published security advisories