Skip to content

Consolidate October dependency updates and clear new npm advisories - #550

Merged
Antnation merged 6 commits into
mainfrom
chore/dependency-refresh-2026-10
Oct 15, 2026
Merged

Antnation merged 6 commits into
mainfrom
chore/dependency-refresh-2026-10

Conversation

@Antnation

Copy link
Copy Markdown
Contributor

Consolidates the open October Dependabot updates into one reviewable change, clears the high-severity npm advisories that now fail npm audit --audit-level=high on main for both lockfiles, and updates rustls for the four RustSec advisories published today, which fail the required audit job on main. Every reviewed inventory the updates invalidate is refreshed in the same PR, and each change is explained below. No advisory ignore, audit level, license allowance, duplicate exception, check or test is added, removed or loosened.

Supersedes #548
Supersedes #547
Supersedes #546
Supersedes #509
Supersedes #544
Supersedes #543
Supersedes #542
Supersedes #539
Supersedes #541
Supersedes #508
Supersedes #510
Supersedes #512
Supersedes #513

What changes

GitHub Actions (#548, #547, #546, #509)

Every commit SHA comes from the matching Dependabot diff, and the tag comments are kept.

Action From To
taiki-e/install-action fa239534 (v2) f7e5d7c9 (v2, 2.87.26)
actions/upload-artifact 043fb46d (v7.0.1) cf430e03 (v7.0.2), all 18 uses
actions/download-artifact 3e5f45b2 (v8.0.1) 9000827c (v8.0.2)
docker/build-push-action 53b7df96 (v7.3.0) c3c9e263 (v7.4.0)

scripts/test_security_mutations_workflows.py pins the exact upload-artifact SHA. #547 left it alone, which would have broken the nightly security-mutations job (it only runs on a schedule, so the PR's CI never showed it). The constant now matches the new pin.

Cargo (#508, #510, #512, #513, plus rustls)

Crate From To Manifest
jsonwebtoken 11.0.0 11.1.0 lock only
syn 3.0.5 3.0.6 lock only
rustix 1.1.4 1.1.5 lock only
jsonschema 0.46.10 0.56.0 ^0.56.0, plus features = ["idna"]
rustls 0.23.45 0.23.46 lock only
  • jsonschema idna feature. jsonschema 0.52 made IDN support an optional, default-on idna feature. This workspace sets default-features = false, so without a change idn-email and idn-hostname would quietly become unknown formats, which the crate accepts. That would weaken validation without any error. The workspace now asks for idna explicitly, which keeps the checks those formats had when idna was mandatory; the lockfile's jsonschema -> idna edge is unchanged. Most of the 0.46 to 0.56 range is correctness fixes, and several make validation stricter: big-integer multipleOf and bounds, uniqueItems treating 0 and -0.0 as equal, Draft 4 integers written with an exponent. It also fixes a stack overflow when compiling unevaluated* beside a $ref that cycles back. New crates: jsonschema-value, strum, strum_macros, heck, all MIT or MIT/Apache-2.0.
  • serde_json float_roundtrip. The new jsonschema-value crate turns on serde_json's float_roundtrip. Feature unification applies it to the whole gateway, so f64 values deserialized from JSON are now correctly rounded (at most one ULP from the old fast path). arbitrary_precision already keeps Value numbers as text, so serialized output does not change.
  • rustls 0.23.46 fixes RUSTSEC-2026-0346/0347/0348/0349, published 2026-10-14. 0349 is the one that affects the gateway: its outbound clients use the default resumption store, which recorded the key-exchange group from an unauthenticated ServerHello. An on-path attacker could use that to make later upstream connections offer only a classical key share, losing post-quantum confidentiality. 0346 needs a custom ClientSessionStore, 0347 needs ECH and 0348 needs the brotli feature, and the gateway uses none of them. The rustls manifest and feature set are otherwise unchanged. I did this in its own commit (109e46c) so it can be split out if another PR lands it first.
  • Lockfile resolution. cargo update -p rustix --precise 1.1.5 (and Dependabot's Bump rustix from 1.1.4 to 1.1.5 #512) also moved ten unrelated windows-sys edges, on errno, socket2, tempfile, cap-*, quinn-udp and others, from 0.61.2 back to 0.52.0 or 0.60.2. The cause is that cargo drops the locks that point at the updated crate's dependencies and then picks the first locked version in range. I put those ten edges back to their exact main values, so rustix's only change is its own version and checksum. cargo metadata --locked and cargo update --workspace --locked accept the lockfile unchanged, and every CI cargo command runs with --locked. Likewise, cargo update for jsonwebtoken and for rustls re-pointed tempfile's getrandom edge from 0.4.3 to 0.3.4. A follow-up cargo update -p getrandom@0.3.4 --precise 0.3.4 returns it to 0.4.3, so that edge is cargo-generated and identical to main.
  • docs/deployment/dependency-controls.md now says the lockfile selects jsonwebtoken 11.1.0. CHANGELOG.md records the refresh and the rustls advisories.

npm

Admin UI (#544, #543, #542): vite 8.3.2, vitest 5.0.3, react-router-dom 7.18.4.

Root (#539, #541): vitest 5.0.3. Wrangler goes to 4.149.0, which supersedes #541's 4.146.0. Wrangler 4.148.0 and earlier pin a Miniflare whose Sharp is older than 0.35.5 (GHSA-wq5f-xc86-pv6w, in the bundled librsvg). 4.149.0 is the lowest release that moves Miniflare (5.20261006.1-alpha) to Sharp 0.35.5 and its libvips 1.3.4 binaries. Wrangler stays an exact pin.

Both lockfiles: source-map-js 1.2.1 -> 1.2.2 (GHSA-68fv-2mgg-jv7q), from a plain npm audit fix with no --force. npm audit now reports 0 vulnerabilities at every severity in both the repo root and admin-ui.

Lockfiles were produced with npm install --package-lock-only --ignore-scripts --before=2026-10-09T00:00:00Z. The --before cutoff means no newly selected transitive version is less than about five days old. For example, rolldown resolves to 1.2.13 rather than 1.2.14, and @cloudflare/workers-types to 5.20261008.1 rather than 5.20261014.1; both of the newer versions were published today.

Inventory changes and why they are safe

cargo-policy-exceptions.json / deny.toml. The duplicate names are the same set (no new group, none removed), and owners and the 2026-12-01 expiry are unchanged. The edits:

  • base64 0.22.1, getrandom 0.2.17 and pem 3.0.6: the parent jsonwebtoken@11.0.0 becomes @11.1.0 (same dependency set). The pem reason now names 11.1.0, in both files.
  • syn: version 3.0.5 becomes 3.0.6, same twelve parents. strum_macros@0.28.0, a derive crate, joins the syn 2.0.119 parents through jsonschema's new strum. The reason ("the remaining derive crates still use syn 2") still holds, and the deny.toml skip (syn@2.0.119) is unchanged.
  • getrandom 0.3.4: jsonschema/referencing 0.46.10 become 0.56.0, and jsonschema-value@0.56.0 is added.
  • hashbrown: referencing moves from 0.16.1 to 0.17.1, so the old-version group shrinks to metrics-util and rsqlite-vfs. The reason now names those parents, in both files.
  • windows-sys 0.61.2: rustix@1.1.4 becomes @1.1.5.

Parent lists were recomputed from cargo metadata --locked --all-features with the same computation validate_graph uses. python scripts/cargo_policy.py config and the graph validation pass locally.

transport-ownership.json. Only graph changes. It was recomputed with transport_guard.py's own graph() from locked metadata, and build_inputs and all 749 syntax scopes are untouched. I reviewed the exposure diff:

  • Package renames for the five bumps. The new crates are listed above, and lazy_static is no longer pulled in by fraction.
  • The workspace dependency record for jsonschema now has req ^0.56.0 and features ["idna"].
  • The serde_json float_roundtrip feature is new, as explained above.
  • hashbrown features move from 0.16.1 to 0.17.1 along with referencing.
  • regex-automata loses its dfa feature, which fancy-regex 0.19 no longer requests (less code compiled).

The rustls entry keeps the same features and dependencies. syn 3.0.6 only changes how expressions are parsed inside None-delimited groups, which come from macro expansion and never from source files the guard reads, so the scope hashes should not move. CI's transport_guard.py check confirms this.

npm-script-policy.json and root allowScripts. Wrangler 4.149.0 changes two locked packages with install hooks. I reviewed them per docs/deployment/npm-lifecycle-policy.md: each tarball was fetched as data, checked against the lockfile SHA-512 and read in memory, with nothing extracted or executed.

  • esbuild 0.28.1 -> 0.28.2: still postinstall: node install.js, and install.js is byte-identical (sha256 612294e2...).
  • workerd 1.20260930.2 -> 1.20261006.1: still postinstall: node install.js, and install.js changes only the embedded release date and version strings used by its binary check and fallback download (sha256 ba1ebba8...).

Both stay allow: false with unchanged reasons; the locked platform optional packages supply the binaries. allowScripts denies the new exact URLs. The admin UI inventory (fsevents only) is unchanged, and the lifecycle doc table names the reviewed versions.

Dropped

Verification

Run locally (Windows, rustc 1.98.0, npm 11.19.1):

  • cargo metadata --locked; cargo update --workspace --locked (no changes); cargo audit --deny warnings passes, where it reports four rustls findings on main
  • python scripts/cargo_policy.py config and the validate_graph stage; 16/16 non-native test_cargo_policy.py tests pass (the native cargo-deny fixtures need the pinned executable, which I did not install locally)
  • test_transport_guard.py: 20/22 pass. The two errors are Windows symlink-privilege fixture setup (WinError 1314), not caused by this change
  • node scripts/npm-script-policy.mjs check for both projects; node --test scripts/test-npm-script-policy.mjs 25/25 pass
  • npm audit --audit-level=high in the root and in admin-ui: 0 vulnerabilities
  • test_security_mutations_workflows.py, test_build_tools.py, test_image_publication.py, test_candidate_scan.py, test_image_evidence.py; scripts/check-supply-chain.py; scripts/build_tools.py check; cargo fmt --check

Left to CI: compiling and testing the gateway, because gateway/build.rs enforces the Node 26.8.2 pin, which this machine does not have. That covers the jsonschema 0.56 API and behaviour, the transport guard's syntax stage, the native cargo-deny policy on Linux and Windows, and the npm marker fixtures on the pinned Node.

Fold in the Dependabot action updates, taking every commit SHA from the
corresponding Dependabot PR diff and keeping the tag comments:

- taiki-e/install-action f7e5d7c9 (v2, 2.87.26), from #548
- actions/upload-artifact cf430e03 (v7.0.2), from #547
- actions/download-artifact 9000827c (v8.0.2), from #546
- docker/build-push-action c3c9e263 (v7.4.0), from #509

The nightly security-mutations workflow contract test pins the exact
upload-artifact SHA it expects. Dependabot did not update it, so #547 on
its own would have broken that scheduled job; update the constant to the
same reviewed SHA.
Fold in Dependabot #508 (jsonwebtoken 11.1.0), #510 (syn 3.0.6), #512
(rustix 1.1.5) and #513 (jsonschema 0.56.0), and refresh the reviewed
Cargo and transport inventories they invalidate.

jsonschema 0.52 turned IDN support into an optional, default-on `idna`
feature. This workspace disables default features, so without a change
the `idn-email` and `idn-hostname` formats would silently become unknown
formats, which the crate accepts. Request `idna` explicitly so they keep
the checks they had when idna was a mandatory dependency; the lockfile
edge to idna is unchanged. The upgrade also brings validation fixes,
including a stack overflow on self-referencing unevaluated* schemas.
New crates: jsonschema-value, strum, strum_macros and heck, all MIT or
MIT/Apache-2.0. No new duplicate crate name appears.

`cargo update -p rustix --precise 1.1.5` also re-pointed ten unrelated
windows-sys edges (errno, socket2, tempfile, and so on) from 0.61.2 to
0.52.0 or 0.60.2: cargo ignores existing locks that point at the updated
crate's dependencies and then picks the first locked version in range.
Those edges are restored to their main values so the only rustix change
is its own version and checksum; `cargo metadata --locked` and
`cargo update --workspace --locked` accept the lockfile unchanged.

cargo-policy-exceptions.json / deny.toml: same duplicate groups, same
owners and expiry. Parent lists follow the version renames, jsonschema
0.56 moves referencing to hashbrown 0.17 (narrowing the 0.16 group),
adds jsonschema-value as a getrandom 0.3 parent and strum_macros as a
syn 2 parent, and the hashbrown/pem reasons now name their actual
parents.

transport-ownership.json: only `graph` changes, recomputed with the
guard's own graph() from locked metadata. Build inputs and syntax
scopes are unchanged; syn 3.0.6 only changes parsing of None-delimited
groups produced by macro expansion, which source files never contain.
Fold in Dependabot #544 (vite 8.3.2), #543 (vitest 5.0.3) and #542
(react-router-dom 7.18.4), and clear the new high-severity source-map-js
advisory GHSA-68fv-2mgg-jv7q (1.0.0 - 1.2.1) with a plain `npm audit fix`
that moves it to 1.2.2. `npm audit --audit-level=high` reports no
vulnerabilities.

The lockfile was produced with `npm install --package-lock-only
--ignore-scripts --before=2026-10-09T00:00:00Z`, so no newly selected
transitive version is less than five days old: rolldown resolves to
1.2.13 rather than 1.2.14, which was published hours before this
change. No lockfile entry gained or lost hasInstallScript, so the admin
UI lifecycle inventory (fsevents only) is unchanged and
`node scripts/npm-script-policy.mjs check admin-ui` passes.
Wrangler 4.148.0 and earlier pin a Miniflare whose Sharp is older than
0.35.5, the fix for GHSA-wq5f-xc86-pv6w. 4.149.0 is the lowest release
that moves Miniflare (5.20261006.1-alpha) to Sharp 0.35.5 and its
patched libvips 1.3.4 binaries, so it supersedes Dependabot's 4.146.0
proposal (#541). Also fold in vitest 5.0.3 (#539) and move source-map-js
to 1.2.2 with a plain `npm audit fix` (GHSA-68fv-2mgg-jv7q). `npm audit
--audit-level=high` reports no vulnerabilities. Resolution used the same
`--before=2026-10-09T00:00:00Z` release-age cutoff as the admin UI.

Wrangler 4.149.0 changes two locked packages that carry install hooks,
reviewed per docs/deployment/npm-lifecycle-policy.md. Both tarballs were
fetched as data, checked against the lockfile SHA-512 and read in memory
without extracting or executing anything:

- esbuild 0.28.1 -> 0.28.2: still `postinstall: node install.js`, and
  install.js is byte-identical (sha256 612294e2...).
- workerd 1.20260930.2 -> 1.20261006.1: still `postinstall: node
  install.js`; install.js changes only the embedded release date and
  version used for its binary check and fallback download
  (sha256 ba1ebba8...).

Neither needs its hook: the locked platform optional dependencies supply
the binaries. npm-script-policy.json records the new identities and
source hashes with unchanged denials and reasons, and allowScripts
denies the new exact tarball URLs. The lifecycle docs name the reviewed
versions and the advisories.
Four rustls advisories were published on 2026-10-14, all fixed in
0.23.46, and `cargo audit --deny warnings` (the required audit job)
fails on main without this update.

RUSTSEC-2026-0349 (GHSA-w44r-93j7-qv28) reaches the gateway: its
outbound TLS clients use the default resumption store, which recorded
the key-exchange group from an unauthenticated ServerHello, so an
on-path attacker could make later connections to that upstream offer
only a classical key share. The other three need a custom
ClientSessionStore (0346), Encrypted Client Hello (0347) or the brotli
feature (0348); the gateway uses none of them.

Lock-only: rustls is transitive, its manifest is unchanged apart from
the version, and the feature set is the same. `cargo update -p rustls`
also re-pointed tempfile's getrandom edge from 0.4.3 to 0.3.4; a
follow-up `cargo update -p getrandom@0.3.4 --precise 0.3.4` returns it
to 0.4.3, so the duplicate inventory is unchanged. The transport graph
is regenerated from locked metadata; build inputs and syntax scopes are
unchanged.
Note the jsonschema 0.56 move and the explicit `idna` feature that keeps
`idn-email`/`idn-hostname` checked, the `float_roundtrip` behaviour that
jsonschema-value brings to serde_json, the lock-only Cargo patch updates,
and the npm updates with the two advisories they clear. Both advisories
are in build and test tooling and do not reach the shipped binary or
the embedded UI bundle.
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 15, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-15T01:20:16.678341Z bf3289e PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@Antnation
Antnation merged commit bf3289e into main Oct 15, 2026
35 checks passed
@Antnation
Antnation deleted the chore/dependency-refresh-2026-10 branch October 15, 2026 01:58
Antnation added a commit that referenced this pull request Oct 15, 2026
Three Vite dev-server advisories published on 2026-10-06 affect 8.3.0
through 8.3.2, which the admin UI locked: GHSA-rq7h-c2jc-7f22 and
GHSA-vfpm-58rq-9qcg let the dev server expose files outside
server.fs.allow, and GHSA-9jrq-w75r-8gcw allows cross-origin script
execution through custom middleware. All three are fixed in 8.3.3;
8.3.4 also matches static aliases on path boundaries and stops adding
the filesystem root to the watcher. They concern `npm run dev` only:
the gateway embeds the production bundle and never runs the dev
server. npm audit's high-severity gate did not catch them because they
are rated medium and low.

The lockfile was produced with the reviewed procedure, `npm install
--package-lock-only --ignore-scripts --before=2026-10-09T00:00:00Z
vite@^8.3.4`, the same cutoff #550 used, so no newly selected version
is younger than that. Vite 8.3.4 was published 2026-10-08; it brings
postcss 8.5.29 and nanoid 3.3.20, both published 2026-10-05. All three
integrity values match the registry and the root lockfile, which
already resolves the same versions through Vitest. The lifecycle
inventory is unchanged (fsevents is still the only flagged location)
and `node scripts/npm-script-policy.mjs check` passes for both
projects.

The lock-only run used --engine-strict=false because this machine's
Node is not the pinned 26.10.0; npm itself is the pinned 11.19.1, and
Node's version does not take part in resolution.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant