Skip to content

v2.68.13

Latest

Choose a tag to compare

@github-actions github-actions released this 13 Nov 03:42
· 25 commits to main since this release

Security

  • klauspost/compress shipped at v1.20.1 with a HIGH zstd decoder panic, and Trivy reported nothing — because the advisory names a subpackage and Go binaries record the module. GO-2026-6865 / CVE-2026-116591 / GHSA-fpmh-frq8-7rj8 (HIGH, CWE-190): a crafted ~27 KB two-block zstd frame makes DecodeAll and the streaming decoder panic with an out-of-range slice index in huff0.ReadTable instead of returning a decode error. Vulnerable <= v1.20.1, patched v1.20.2. The published applegater/caddyui-caddy:stable shipped v1.20.1. Dockerfile.caddy now pins --with github.com/klauspost/compress@v1.20.4.
    • This is a third, distinct way a finding can be invisible to Trivy, and the first two explanations do not apply. Unlike the go-jose and crowdsec cases ([[T-0646]]), this advisory has a CVE id and is in OSV — /v1/vulns/ answers for all three ids. The cause is a package-vs-module naming mismatch: OSV and GitHub name the affected package github.com/klauspost/compress/zstd, while a Go binary's build info and any go.mod record the module github.com/klauspost/compress. Trivy matches on the recorded name, finds no advisory under the module path, and reports nothing — while listing the package at v1.20.1 in its own Packages output, so this is not a DB-lag miss either (advisory published 2026-10-27, Trivy DB 2026-11-12). Measured, not inferred: a probe go.mod requiring github.com/klauspost/compress v1.20.1 yields 0 findings; one requiring github.com/klauspost/compress/zstd v1.20.1 yields the HIGH with FixedVersion: 1.20.2 — same Trivy, same DB, same version. govulncheck is unaffected because the Go advisory names the module correctly and matches on symbols.
    • The image gate therefore reads unchanged, not improved, and that is correct rather than a failure. Gate 1 is set arithmetic over Trivy findings; a finding Trivy never reported cannot be "closed" by fixing it. Trivy's counts are 4 before and 4 after, with the klauspost version visibly moving v1.20.1 → v1.20.4 between the two scans. The fix is proven by govulncheck -mode=binary on the extracted binaries instead: 2 findings → 1, GO-2026-6865 closed, 0 introduced. This is the gate's documented "what this verdict is not evidence of" limit meeting a concrete case.
    • Reachability: linked and statically reachable, but no live code path decodes hostile input — and the fix ships anyway. Re-derived against the artifact, which mattered: the build runs -tags nobadger,nomysql,nopgx, so a first probe that omitted the real build tags and most plugins produced a confident but wrong badger/smallstep-nosql route — dgraph-io/badger is not linked into the binary at all (0 strings). A probe rebuilt with all 13 plugins, modules/standard and the real tags reproduces the binary's klauspost package set exactly. go mod why without -vendor is also misleading here: it reports a prometheus promhttp route that runs through prometheus/client_golang/prometheus.test and is test-only. The only non-test importers in the shipped closure are Caddy's own modules/caddyhttp/encode/zstd and DeRuina/timberjack (log rotation). The advisory is a decoder bug; Caddy's encode/zstd is encoder-only — it calls zstd.NewWriter and nothing else, and zstd_precompressed.go merely declares the .zst suffix so the file server can serve pre-compressed files untouched. Caddy compresses responses; it never decompresses a zstd request body. A sweep of the entire 1274-package shipped closure for non-test zstd.NewReader / .DecodeAll( finds hits only inside klauspost/compress/zstd itself, with a negative control confirming the same sweep does fire for zstd.NewWriter. govulncheck calls the symbols reachable because they are present and linked, which means linked, not exploitable. Bumped regardless: the fix is free, the module set is otherwise untouched, and a future config or plugin that decodes zstd would make it live.
    • The pin is load-bearing, not slack. MVS resolves to exactly v1.20.1 on its own — caddy 2.11.9, caddy-l4 v0.1.3, caddy-crowdsec-bouncer v0.15.0 and prometheus/client_golang v1.25.0 each require v1.20.1 and nothing in the graph requires higher — so without this line the build stays on the vulnerable version. Unlike the crowdsec floor, deleting it regresses the image, which the build now asserts: go version -m is written to a file and checked for >= v1.20.4, and a build with the pin removed and the assertion kept fails with the assertion's own message showing v1.20.1, verified rather than assumed. The pattern is deliberately narrow — it accepts only v1.20.x for x >= 4, with a trailing whitespace class so a pre-release pseudo-version like v1.20.4-0.2026…, which sorts below v1.20.4, cannot satisfy it — so a future bump to v1.21.0 fails the build until the pattern is widened in the same commit. That fails closed by design. v1.20.4 rather than the minimum v1.20.2 because v1.20.2/.3/.4 are all clean at both the module and zstd subpackage paths in OSV, the per-repository GitHub endpoint lists only two advisories ever (both fixed at or below v1.20.2), and v1.20.4 is the current release.
    • The weekly pin watcher was not blind to this, and the gap was elsewhere. scripts/check-pinned-modules.sh's image sweep ([[T-0669]]) queries GitHub's per-repository endpoint, which is keyed on the repository, so the subpackage spelling never arises. Replaying its exact matcher against the live payload matches GHSA-fpmh-frq8-7rj8 via the non-module-path branch, and sourcing its real range_excludes_version / patch_already_in_version confirms the row is kept at v1.20.1 and excluded at v1.20.4. The actual gap was that klauspost/compress had no --with pin, so it could only ever appear through the graph sweep — which is capped (40 unauthenticated, 400 with a token) and reports rather than gates. Pinning it moves the module into the per-pin section that is queried every run.
    • Caddy's module set is byte-identical across the change (caddy list-modules --versions diff empty), the binary module diff is the single klauspost/compress line, caddy version stays v2.11.9, and getcap still reports cap_net_bind_service=ep.

Rebuilt and published: applegater/caddyui-caddy:v2.68.13, :stable, :latest.

Applying it needs a container recreate, not just a tag move. On each host, use Portainer's Re-pull image and recreate — measured hosts split between stale-image and stale-container failure modes. Verify from inside the container (docker exec <c> sh -c 'caddy version'), never from the tag.

Rollback: applegater/caddyui-caddy:v2.68.12.