Repository navigation
SECURITY: microvm-rootfs: libexpat 2.10.0-r0 (4 CVEs), zlib 1.4-r0 override (CVE-2026-85091, CVE-2026-4738) - #1380
Conversation
alpine-pins flags libexpat-2.9.0-r0 as missing fixes for CVE-2026-113351, CVE-2026-115506, CVE-2026-117610 and CVE-2026-118370, all fixed in 2.10.0-r0. c-ares 1.34.8-r0 -> 1.34.9-r0 is the only other stale pin in the closure (no recorded security fix) and rides along. Both new builds depend only on so:libc.musl in the v3.24 APKINDEX, so nothing else in the closure moves. sha256s computed from dl-cdn.alpinelinux.org downloads whose size and control checksum match the APKINDEX. The four objects are not yet mirrored to gs://minimal-staging-archives. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
…s); keep c-ares alpine-minirootfs-3.24.1 ships zlib 1.3.2-r0, which the v3.24 secdb lists as missing CVE-2026-85091 (fixed 1.3.2-r1) and CVE-2026-4738 (fixed 1.4-r0). alpine-pins only reads the apk pins, so it never reported it. git and libcurl link so:libz.so.1, so the guest loads it. Overlay zlib-1.4-r0 in the base-layer override group, next to musl / libssl3 / ca-certificates-bundle. build.sh unpacks the minirootfs first and overlays every .apk after it, replacing the libz.so.1 symlink, and its apk-db reconcile moves the recorded V: to 1.4-r0. zlib-1.4-r0 depends only on so:libc.musl. sha256s are from dl-cdn downloads. Revert c-ares to 1.34.8-r0: 1.34.9-r0 carries no secfix, nothing in the APKINDEX requires it (libcurl depends on unversioned so:libcares.so.2), and 1.34.8-r0 is already mirrored, so this drops two objects from the mirror list. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
libexpat-2.10.0-r0 and zlib-1.4-r0 (aarch64 + x86_64) are now on gs://minimal-staging-archives with the sha256s in build.ncl. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
agent-137
left a comment
There was a problem hiding this comment.
Review on Karl's behalf. This is a comment, not an approval. Nothing was built for it. The findings come from reading the declarations and fetching upstream sources, and each one says how it was checked. The CI rollup on this head is now SUCCESS.
CI: 15/15 green, CodeRabbit SUCCESS ("no actionable comments"), no human reviews, not a draft. The PR body itself says CI does not fetch the gs:// sources, so the green run does not show that the image builds.
Verified clean:
- Hashes: all four sha256s match fresh dl-cdn downloads. The staging mirror now serves all four objects with the same sha256s. The argument order (arm64 first, then amd64) matches the
apkhelper. - CVE claims: they agree with
secdb.alpinelinux.org/v3.24/main.json.- expat
2.10.0-r0fixes CVE-2026-113351, -115506, -117610 and -118370. - zlib
1.3.2-r1fixes CVE-2026-85091, and1.4-r0fixes CVE-2026-4738. - c-ares
1.34.8-r0is the newest secfix entry, so keeping 1.34.8 is right.
- expat
- CVE IDs: none is malformed. CVE-2026-4738 has a low sequence number for November, but the format is valid and it appears in the secdb.
- Base layer: the mirrored
alpine-minirootfs-3.24.1-x86_64.tar.gzmatches its pinned sha256, and its db recordszlib V:1.3.2-r0, as the PR says. The zlib-1.4-r0 apk ships onlyusr/lib/libz.so.1andusr/lib/libz.so.1.4, so the file rename the PR describes is real.
Should-fix
-
packages/microvm-rootfs/build.ncl:248-256and:322-336: a release bump exists, and the comment says it does not. The comment says "the newest minirootfs RELEASE is still 3.24.1 … There is no 3.24.2 to move to". Alpinev3.24/releases/x86_64/latest-releases.yamlnow lists 3.24.4. Thelib/apk/db/installedinalpine-minirootfs-3.24.4-x86_64.tar.gzalready records:ca-certificates-bundle 20261020-r0libcrypto3/libssl3 3.5.9-r0musl/musl-utils 1.2.6-r3zlib 1.3.2-r1
So every existing override is redundant at 3.24.4. zlib 1.4-r0 still needs an override there, because CVE-2026-4738 is not fixed in 1.3.2-r1. The zlib override in this PR is justified, and the libexpat change is a proper closure upgrade. The right fix overall is to move
alpine_releaseto 3.24.4 and keep only the zlib override. The PR body's line "might carry the zlib fix" is half right. Fix the stale comment in this PR, or open a tracked follow-up, so the next reader does not trust "no 3.24.2". Verified by downloading the 3.24.4 x86_64 minirootfs from dl-cdn. aarch64 was not checked. -
build.ncl:336andbuild.sh:128-160: the override has no expiry check, so it can stay forever. "DROP THIS once…" is only a comment. The reconcile rewritesV:to whatever the overlay installed, and the fail-closed loop only checks that the rewrite landed. It never compares that version with the base layer's version. Failure scenario: someone bumpsalpine_releaseto a minirootfs that ships zlib 1.4-r1 or later and forgets this entry. The overlay then silently downgrades libz. The db is rewritten to1.4-r0, and the build passes. The openssl comment at:265-267names exactly this risk ("pinned BELOW its base layer"). The pattern predates this PR, but this PR adds another instance of it. Suggested fix: record the expected base version for each override (for examplezlib 1.3.2-r0), and fail the build when the base db records anything else. That turns the DROP comment into a build failure. A tracking issue is an acceptable alternative. Verified by readingbuild.shat PR head. -
build.ncl:329-334: the old vulnerable libz stays on disk, and the db file list goes stale. The overlay leavesusr/lib/libz.so.1.3.2on disk. That is the 1.3.2-r0 binary without the CVE-2026-85091 fix. After the reconcile, the db stanza readsP:zlib V:1.4-r0but still listsR:libz.so.1.3.2and notlibz.so.1.4. Failure scenario: a file-level scanner (a binary cataloger) finds libz 1.3.2 in the image and reports CVE-2026-85091 and CVE-2026-4738 on an image whose fix is exactly those two CVEs. That leads to repeat triage or an alert people learn to ignore. The PR discloses this and defers it. The fix is one line inbuild.sh(removeusr/lib/libz.so.1.3.2after the overlay, guarded on the zlib override). That is cheap enough to do in this PR, or track it in a filed issue. Verified from the base db stanza and the apk file listing.
Nit
- PR body is stale. "Draft until these four objects are mirrored … a public HEAD returns 404 for all four" is no longer true. The mirror serves all four with matching hashes (I checked), and the PR is not a draft. Update the body before merge so the record is accurate.
Verdict #1380: the pins and CVE claims are correct and verified. Mergeable as a fast security fix, once the stale "no 3.24.2" comment is fixed or tracked, and with follow-ups for the override expiry check and the orphaned libz.so.1.3.2. The build itself is unverified.
….so.1.3.2 Review (agent-137 on #1380): - build.sh records the base-layer version each override replaces (OVERRIDE_BASE, read from the 3.24.1 minirootfs db on both arches) and fails when the minirootfs records anything else, or when an overlay replaces a base package with no line. A DROP THIS comment becomes a build failure instead of a silent downgrade below a newer base. - After the overlay, only the file libz.so.1 points at is kept: the base layer's libz.so.1.3.2 is removed and zlib's db entry moves to libz.so.1.4, so a file scanner does not report the fixed CVEs. - The comment no longer says there is no newer minirootfs: 3.24.4 exists and retires every override but zlib 1.4-r0 (a follow-up). Exercised the new build.sh sections against the real 3.24.1 x86_64 minirootfs and zlib-1.4-r0 apk: positive run reconciles and removes the orphan; base zlib moved to 1.4-r1 fails; an untabled base override fails. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
norrietaylor
left a comment
There was a problem hiding this comment.
Scope matches the title and the update note. The diff bumps the libexpat closure pin, adds a zlib base-layer override, adds the OVERRIDE_BASE expiry check, and adds the orphaned libz.so.1.3.2 cleanup. Clobber none. git log on packages/microvm-rootfs shows the recent musl, ca-certificates-bundle, brotli and libcurl security commits, and this diff keeps all of them. Correctness The OVERRIDE_BASE table matches every base-layer override in build.ncl: libcrypto3, libssl3, ca-certificates-bundle, musl, musl-utils and zlib. No closure pin replaces any other minirootfs package. The two checks run after the -s "$DB" guard and before the reconcile, so they read the base-layer V: values. The reverse check makes the build fail when a new override has no table line. In the orphan cleanup, the libz.so.1.* glob always matches the live file, so it never runs on a literal pattern. A relative readlink target resolves correctly, and an absolute one fails closed. The zlib stanza rewrite only touches R: lines in the zlib stanza. The new sha256 values match the values in the PR body. Not checked The image build and the gs:// fetch of the new objects. Both need a build.
| # 3.24.1, so bumping `alpine_release` could not pick the fix up. 3.24.4 has | ||
| # since been released and records libcrypto3/libssl3 3.5.9-r0, musl 1.2.6-r3, | ||
| # ca-certificates-bundle 20261020-r0 and zlib 1.3.2-r1, so moving to it | ||
| # retires every override here except zlib 1.4-r0 (tracked separately). |
There was a problem hiding this comment.
nit · The comment says zlib 1.4-r0 is "tracked separately", but it names no issue. The PR body says that the 3.24.4 move is a later PR. Link the issue that tracks the move here. Then the next reader can find out when this block and the OVERRIDE_BASE lines can go.
Two security fixes to the microvm-rootfs guest image:
2.9.0-r0->2.10.0-r0(closure pin). alpine-pins (nightly 2026-11-04 through 11-06) reports CVE-2026-113351, CVE-2026-115506, CVE-2026-117610 and CVE-2026-118370, all fixed in2.10.0-r0per the v3.24 secdb.1.3.2-r0->1.4-r0(new base-layer override). The guest's zlib comes fromalpine-minirootfs-3.24.1(itslib/apk/db/installedrecordszlib 1.3.2-r0), not from the closure. The v3.24 secdb lists CVE-2026-85091 as fixed in1.3.2-r1and CVE-2026-4738 as fixed in1.4-r0. git and libcurl linkso:libz.so.1, so the guest loads it. alpine-pins only reads the apk pins, never the minirootfs db, so it did not report this.zlib goes in the override group next to musl / libssl3 / ca-certificates-bundle. build.sh unpacks the minirootfs first and then overlays every
.apkwithtar -xof, so the overlay replacesusr/lib/libz.so.1; the apk-db reconcile then rewrites zlib's recordedV:to1.4-r0(and fails the build if it doesn't land). In the v3.24 APKINDEX,libexpat-2.10.0-r0andzlib-1.4-r0depend only onso:libc.musl-<arch>.so.1, and sonames are unchanged (so:libexpat.so.1,so:libz.so.1), so nothing else moves.zlib's real file is renamed between versions (
libz.so.1.3.2->libz.so.1.4). The overlay replaces thelibz.so.1symlink, so the loader gets 1.4, but the oldusr/lib/libz.so.1.3.2file stays on disk, unreferenced, because build.sh only adds and replaces paths. A file-level scanner may still flag it (a false positive, not a missed CVE). Removing base-layer files an override no longer ships would be a build.sh change; I left it out of this PR.c-ares stays at
1.34.8-r0. An earlier revision of this PR bumped it to1.34.9-r0, but that build has no secfix, nothing in the APKINDEX needs it (libcurl depends on unversionedso:libcares.so.2), and 1.34.8-r0 is already mirrored.Alternative not taken: moving
alpine_releaseto 3.24.4 (now the latest release) might carry the zlib fix in the base layer and let some overrides go. It's a bigger change, and its minirootfs isn't mirrored either.Mirrored
All four objects are on the staging mirror and serve the sha256s below (re-checked 11-09; agent-137 checked them independently).
alpine/v3.24/main/aarch64/libexpat-2.10.0-r0.apk99144fe0b1bb9316a3bc30c38707996767806661f4a0456fcf4565ac92e79b0falpine/v3.24/main/x86_64/libexpat-2.10.0-r0.apke7e7408ea787ed06d00159b0ff1f9b65a63cf634f36d9e0fa7512ce2c489a503alpine/v3.24/main/aarch64/zlib-1.4-r0.apkd3e30d012a3445222b0b4a28f1d4d97ce67387f8b48561921111696da174f0dealpine/v3.24/main/x86_64/zlib-1.4-r0.apk7b05bc778b8b25cd880835b40114182bd4d1581717504d0129c810c0a14291cbHow the sha256s were checked: each was computed from a dl-cdn download, and an independent reviewer got the same values from their own download. The libexpat files' sizes match the APKINDEX
S:field and their control-segment SHA-1 matchesC:. I didn't verify the APKINDEX's own Alpine signature. Nothing was built locally, and CI doesn't fetch the gs:// sources, so a green CI run doesn't show the image builds.nickel formatis clean.🤖 Generated with Claude Code
Summary by CodeRabbit
Update: review follow-ups (bb3b2a4)
OVERRIDE_BASErecords the base-layer version each override replaces (read from the 3.24.1 minirootfs db, both arches: libcrypto3/libssl3 3.5.7-r0, ca-certificates-bundle 20260611-r0, musl/musl-utils 1.2.6-r2, zlib 1.3.2-r0). The build fails when the minirootfs records anything else, or when an overlay replaces a base package that has no line. So movingalpine_releaseforces every override to be re-decided, instead of silently pinning the image below a newer base.libz.so.1.3.2is removed after the overlay, and zlib's db file entry moves tolibz.so.1.4. The paragraph above about it staying on disk no longer applies.Exercised the new build.sh sections against the real 3.24.1 x86_64 minirootfs and the zlib-1.4-r0 apk: the positive run reconciles and removes the orphan; a base layer with zlib 1.4-r1 fails; an override of a base package with no table line fails. The full image is proved by the build bot's run on this head.
🤖 Generated with Claude Code