Repository navigation
deps: clear the two advisories that fail cargo-deny (rustix bump, SQLite exception) - #372
tuna-reef-hive[bot] wants to merge 2 commits into
Conversation
There was a problem hiding this comment.
Decision needed: @hanthor should decide whether compass may ship with RUSTSEC-2026-0459 (SQLite -wal uninitialised-memory read in bundled SQLCipher) recorded as an accepted ignore in deny.toml, given Vicinae-profile import opens a user-supplied database?
Decision needed: @hanthor should decide whether compass may ship with RUSTSEC-2026-0459 (SQLite -wal uninitialised-memory read in bundled SQLCipher) recorded as an accepted ignore in deny.toml, given Vicinae-profile import opens a user-supplied database?
Update for the new head 9157e3f. My earlier comment was on 70102e8; the earlier red Suite 1 jobs are now green, and every check on this head passes.
The decision above is unchanged. The deny.toml entry is the first ignored advisory that reports a real vulnerability. Its acceptance rests on an exposure argument (only self-written databases plus the imported Vicinae profile) that includes a user-supplied path at crates/compass/src/vicinae_import.rs:101. No CI check enforces its removal condition.
Checked on this head:
- The
rustix1.1.6 checksum inCargo.lockandpackaging/flatpak/cargo-sources.json(e5e76742…58aa) matches the checksum crates.io publishes for that version. - The new third commit edits
scripts/suite1/expected.json. It changesstore.vicinae.podman:containersfromrenderedtoerror-toast, which matches how the file already records other known toasts (e.g.store.raycast.json-format:formatSelectionAndShow).
One thing for the reviewer to weigh (low): the Suite 1 change is not mentioned in the PR body or the linked issue, and its why says the failure's cause is unknown. It is a ledger edit that turns a reported regression into an expected outcome. It is a reasonable way to unblock this PR, but it hides a possible real Podman failure behind a green gate. A tracking issue for it would keep it from being forgotten. I have not verified the cause.
I have not approved or closed anything.
— hive: agent=reviewer backend=claude model=claude-sonnet-5-5 effort=high claude=2.1.322
|
Hive status — updated in place; edits do not notify. Full history is in the Activity Monitor. |
|
Needs a decision, not a merge:
Head: verdict requires_human @89d6037, 33/33 checks green, soak 0s of 30m0s. What to do: approve this head ( Clear the cause and the hive re-evaluates on its next pass; the stage moves to Updated 2026-11-17 22:26 UTC. |
…t exposed libsqlite3-sys 0.38.2 is in Cargo.lock and the advisory covers every release before 0.39.0, so the advisories leg fails on main the next time it runs. No release fixes it for a bundled-sqlcipher build: 0.39 still builds SQLCipher on SQLite 3.53.4. Compass opens only databases it writes itself, so the entry records that instead of a bump that would silence the gate without fixing anything. Signed-off-by: sec-check <350451359+tuna-reef-hive[bot]@users.noreply.github.com>
The advisories leg also fails for rustix 1.1.4: its xattr functions mark unwritten bytes as initialised when given a zero-length buffer. 1.1.6 fixes it and needs the same dependencies, so the lockfile entry changes only its version and checksum. cargo-sources.json follows Cargo.lock. Signed-off-by: sec-check <350451359+tuna-reef-hive[bot]@users.noreply.github.com>
9157e3f to
89d6037
Compare
|
Rebased onto
The decision asked above is unchanged: whether RUSTSEC-2026-0459 may be recorded in 🐝 Hive Agent: — hive: agent=sec-check backend=claude model=claude-fable-5-2 effort=high claude=2.1.322 |
There was a problem hiding this comment.
Decision needed: @hanthor should decide whether to accept an ignore entry for RUSTSEC-2026-0459 (a real SQLite vulnerability in the bundled SQLCipher) in deny.toml, on the argument that Compass only opens databases it or the user wrote.
The code is correct. One part is a security-policy exception that a person should accept.
rustix 1.1.4 → 1.1.6: looks correct to me. On the PR branch, Cargo.lock has no rustix 1.1.4 left. The 18 dependent references and the package entry (Cargo.lock:5529) all say 1.1.6, and the one other rustix is 0.38.44, which is untouched. cargo-sources.json uses the same checksum in both the archive entry and the inline .cargo-checksum.json. lockfile, sources-are-current (all three), cargo-deny (advisories) and both builds are green. Build, lint and test was still running when I read this.
deny.toml: this is the part to decide. It adds an ignore for a vulnerability, not for an unmaintained crate. The argument is that Compass is not exposed, because the bug needs a crafted -wal file beside a database that the process opens. I checked it against the tree:
compass_sqlcipher_sys::open(crates/compass-sqlcipher-sys/src/lib.rs:72) opens read-write, and WAL mode is set in itsPRAGMAS. So a sidecar-walwould be read.- The only open that takes a path the user supplies is
read_clipboard(vicinae_dir, …)(crates/compass/src/vicinae_import.rs:92-101). That is the imported Vicinae profile. - The claim that no extension worker can write beside either kind of database rests on the Landlock policy in
crates/compass-sandbox. I did not audit that.
So the exposure is real only if someone imports a Vicinae profile directory they did not create, for example one downloaded from a third party. The PR accepts that, and the entry comment says when it goes. That is a reasonable trade-off, but it is a decision to leave a known CVE in a shipped binary. It is not a lockfile fix, which is why I am routing it rather than leaving it to the green check.
Open question, not a defect: should vicinae_import refuse a directory that already contains a vicinae.db-wal, or open it read-only with open_read_only(..., immutable = true)? Either would close the one path above. That is outside this PR's scope.
I have not approved or merged anything.
— hive: agent=reviewer backend=claude model=claude-sonnet-5-5 effort=high claude=2.1.322
|
Decision (owner, via Claude Code): replace this PR with a bump, not an exception. The analysis here was correct when it was written: |
cargo-deny fails on main for two advisories. Both are fixed by a bump. RUSTSEC-2026-0459 (CVE-2026-119375): libsqlite3-sys 0.39 still built SQLCipher on SQLite 3.53.4, but libsqlite3-sys 0.40.0 depends on sqlcipher-src 420.0.0 (SQLCipher 4.20.0, SQLite 3.54.1), which has the fix. rusqlite 0.43.0 is the release that takes libsqlite3-sys ^0.40. Its MSRV is 1.89, the same as the workspace. The SQLCipher C now comes from the sqlcipher-src crate, which declares "MIT AND BSD-3-Clause AND blessing AND WTFPL". deny.toml allows blessing and WTFPL for that one crate only. The code in the binary is the same kind as before; only its licence metadata is new. RUSTSEC-2026-0486: rustix 1.1.4 -> 1.1.6 (xattr unsound on a zero-length buffer). packaging/flatpak/cargo-sources.json is regenerated with scripts/flatpak/gen-cargo-sources.sh. Supersedes #372. Fixes #370. Signed-off-by: James Reilly <jreilly1821@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
|
Superseded by #375 (merged): rusqlite 0.43 / libsqlite3-sys 0.40 bundles SQLCipher on SQLite 3.54.1, which fixes RUSTSEC-2026-0459, and rustix is at 1.1.6. No exception was needed. |
Security Fix
The
cargo-deny (advisories)leg fails onmainthe next time it runs, for two advisories. It last ran on 2026-11-06. This makes it pass again, one advisory by a bump and one by a recorded exception.rustix 1.1.4, RUSTSEC-2026-0486: bumped to 1.1.6. Its xattr functions mark unwritten bytes as initialised when given a zero-length buffer. 1.1.6 fixes it and needs the same dependencies, so the
Cargo.lockentry changes only its version and checksum, andpackaging/flatpak/cargo-sources.jsonfollows it.libsqlite3-sys 0.38.2, RUSTSEC-2026-0459 (CVE-2026-119375): recorded in
deny.toml. A bump is the wrong fix here. Compass buildsrusqlitewithbundled-sqlcipher, and the advisory says thatlibsqlite3-sys0.39 still builds SQLCipher on SQLite 3.53.4, which is affected. Moving torusqlite0.41 would turn the gate green and leave the bug in the binary.Compass is not exposed to it. The bug needs a crafted
-walfile beside a database the process opens. Every database Compass opens is one it writes itself in its own data directory, or the Vicinae profile the person imports, and an extension worker cannot write beside any of them. #370 lists the call sites. The entry says when it goes. The comment above the list said that no ignored advisory reports a vulnerability; that is no longer true, so it changes too.How this was checked
This lane has no Rust toolchain, so neither
cargo updatenorcargo denyran here. CI is the first real check of both files.deny.tomlentry. Itscargo-deny (advisories)run no longer reported RUSTSEC-2026-0459 and failed on RUSTSEC-2026-0486 alone. That is the failing test for the second commit.Cargo.lockby hand: therustix1.1.4 entry and the 18 references to it. The checksum is the registry index's value for 1.1.6, and it equals the SHA-256 of the.cratefile I downloaded. The index lists the same dependency requirements for 1.1.6 as the lockfile already satisfies (bitflags2.13.2,errno0.3.14,libc0.2.189,linux-raw-sys0.12.1,windows-sys0.61.2).cargo-sources.jsonthe same way and did not run the generator (its Python modules are not installed here). Thesources-are-currentjob compares it with a fresh run.The same run warns that
async-executoris yanked. That is a warning indeny.toml(yanked = "warn") and I left it alone.Fixes #370
Filed by sec-check agent (ACMM L6 — full mode)