Skip to content

security: x/net v0.62.0, the Go floor at 1.27.3, and a lane that would have said so - #26

Merged
tannevaled merged 1 commit into
mainfrom
raise-x-net-to-0.62.0
Oct 31, 2026
Merged

tannevaled merged 1 commit into
mainfrom
raise-x-net-to-0.62.0

Conversation

@tannevaled

Copy link
Copy Markdown
Contributor

Two reachable advisories, and nothing here could have reported either

This repository had no vulnerability lane at all. Found 2026-10-31 by a sweep of the 58 Go repositories across go-webengine, go-compressions, go-filesystems and go-images — the first anyone had run over them. 58 scanned, 0 unread, 2 reaching something: this repository and go-webengine/engine, both the same x/net advisory.

advisory fixed in reachable through
GO-2026-6860 golang.org/x/net v0.62.0 session.go:106:29 engineRender → Engine.RenderWithLinks → html.Parse
cmd/browserproxy/main.go:90:29 run → http.Server.Shutdown → html.ParseFragment
GO-2026-6809 crypto/x509 Certificate.Verify go1.26.10 / go1.27.3 cmd/browserproxy/main.go:75:13, because go.mod declared go 1.27.2 and the standard library is judged by the version the module claims

Those call paths are the actual job of this program — it parses pages it did not write. Not an init artefact.

⛔ Renovate could not have found the first one. x/net was an indirect requirement (// indirect), so it is not a floor this module declares for itself and no manager had a version string to compare or rewrite. Nothing watched it. The same advisory sat green for days in two go-pkgx repositories three days earlier and was found there by pull requests that touched nothing related.

Measured, one variable

Same tree, same govulncheck, GOOS=linux GOARCH=amd64:

before   2 vulnerabilities, both with the symbol traces above
after    No vulnerabilities found.   Your code is affected by 0 vulnerabilities.

The floor is named in three places, and all three move

Raising go.mod alone would leave the lane building the vulnerable toolchain and the README advertising a floor two patches below the real one:

go.mod go 1.27.2 → 1.27.3
ci.yml GO_VERSION 1.27.2 → 1.27.3 (one definition, three jobs)
README badge Go 1.27.1+ → Go 1.27.3+ — it was already stale by two patches

And the hole that hid them is closed

.github/workflows/vuln.yml: govulncheck judged by vulngate rather than by an exit status, on push, pull_request, a daily cron and workflow_dispatch. A gate that only fires on a push has the coverage of the repository's activity, and the quietest repository needs it most.

Four choices in it, each with a reason rather than a habit:

  • ⛔ Its own workflow, not a job in ci.yml. A schedule there would run the wasm end-to-end lane and the cross-build matrix every night to ask a question about dependencies.
  • ⛔ The toolchain comes from go.mod, not from ci.yml's GO_VERSION. The declared floor is what a consumer running go install with GOTOOLCHAIN=local honours, so it is the version this question is about — and reading it there keeps one source of truth instead of a second copy that rots out of step.
  • ⛔ The judge is @latest, not a pin. A pin inside a run: block is read by no dependency manager, so it rots in silence: go-gitsafe/gitsafe pins [email protected] in its own copy of this step while go-fleettools/fleettools is at v0.51.0 — thirty-one minor versions, and nothing said so.
  • ⛔ A disabled cron reports nothing rather than failing (GitHub switches scheduled workflows off after 60 days of inactivity), hence workflow_dispatch, a comment saying silence is not green, and a badge so the lane is visible rather than merely present.

It ships green, and that was checked rather than hoped

The gate was run by hand before being committed: govulncheck wrote 883 KB of JSON and vulngate exited 0, judging GO-2026-5932 a notice (required, not called, no released fix published) with the advisory database 4 days old against its 21-day cap.

gofmt clean · go build ./... ok · go vet ./... silent · go test ./... 2 packages ok · actionlint silent.

Not in this PR

go-webengine/engine reaches the same x/net advisory and its nested bench module is unreadable to govulncheck (go: updates to go.mod needed — Renovate updates the parent and leaves the nested module behind). That repository is claimed by another session; the finding has been sent to them rather than acted on here.

🤖 Generated with Claude Code

…d have said so

TWO REACHABLE ADVISORIES, and nothing in this repository could ever have
reported either: it had no vulnerability lane at all.

Found on 2026-10-31 by a sweep of the 58 Go repositories across
go-webengine, go-compressions, go-filesystems and go-images -- the first
anyone had run over them. 58 scanned, 0 unread, 2 reaching something:
this repository and go-webengine/engine, both the same x/net advisory.

  GO-2026-6860  golang.org/x/net v0.61.0 -> v0.62.0
    session.go:106:29              engineRender -> Engine.RenderWithLinks
                                   -> html.Parse
    cmd/browserproxy/main.go:90:29 run -> http.Server.Shutdown
                                   -> html.ParseFragment

  GO-2026-6809  crypto/x509 Certificate.Verify, fixed in go1.26.10 and
                go1.27.3; reachable because go.mod declared `go 1.27.2`
                and the standard library is judged by the version the
                module claims.

Those call paths are the actual job of this program: it parses pages it
did not write. This is not an init artefact.

⛔ RENOVATE COULD NOT HAVE FOUND THE FIRST ONE. x/net was an INDIRECT
requirement (`// indirect`), so it is not a floor this module declares
for itself and no manager had a version string to compare or rewrite.
Nothing watched it. The same advisory sat green for days in two go-pkgx
repositories three days earlier and was found there by pull requests
that touched nothing related.

Measured, same tree, same govulncheck, GOOS=linux GOARCH=amd64:

  before   2 vulnerabilities, both with symbol traces (above)
  after    No vulnerabilities found. 0 affected.

THE FLOOR IS NAMED IN THREE PLACES AND ALL THREE MOVED, because raising
go.mod alone would have left the lane building the vulnerable toolchain
and the README advertising a floor two patches below the real one:

  go.mod            go 1.27.2 -> 1.27.3
  ci.yml            GO_VERSION 1.27.2 -> 1.27.3 (one definition, 3 jobs)
  README badge      "Go 1.27.1+" -> "Go 1.27.3+" (already stale by two)

AND THE HOLE THAT HID THEM IS CLOSED: .github/workflows/vuln.yml, a
govulncheck judged by go-fleettools/vulngate rather than by an exit
status, on push, pull_request, a DAILY cron and workflow_dispatch.

⛔ Its own workflow and not a job in ci.yml: a schedule there would run
the wasm end-to-end lane and the cross-build matrix every night to ask a
question about dependencies.

⛔ It takes the toolchain from go.mod and not from ci.yml's GO_VERSION.
The declared floor is what a consumer running `go install` with
GOTOOLCHAIN=local honours, so it is the version the question is about --
and reading it there keeps one source of truth instead of a second copy
that rots out of step.

⛔ The judge is @latest and not a pin, and the reason is measured: a pin
inside a `run:` block is read by no dependency manager, so it rots in
silence. go-gitsafe/gitsafe pins [email protected] in its own copy of
this step while go-fleettools/fleettools is at v0.51.0 -- thirty-one
minor versions, and nothing said so.

⛔ A disabled cron reports nothing rather than failing: GitHub switches
scheduled workflows off after 60 days without repository activity. Hence
workflow_dispatch, a comment saying silence is not green, and a badge in
the README so the lane is visible rather than merely present.

The gate was run by hand before being committed, which is how I know it
ships green: govulncheck wrote 883 KB of JSON, vulngate exited 0 judging
GO-2026-5932 a notice (required, not called, no released fix published)
with the advisory database 4 days old against its 21-day cap.

gofmt clean, go build ./... ok, go vet ./... silent, go test ./... 2
packages ok, actionlint silent.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
@tannevaled
tannevaled merged commit 957dd29 into main Oct 31, 2026
9 checks passed
@tannevaled
tannevaled deleted the raise-x-net-to-0.62.0 branch October 31, 2026 20:54
tannevaled added a commit to go-webengine/engine that referenced this pull request Oct 31, 2026
…ne (#331)

A peer session's vulnscan sweep of 58 Go repositories (the first run
over this population) flagged go-webengine/engine as CALLED with a
reachable advisory, and a browserproxy#26 precedent to copy. Verified
independently with govulncheck before acting:

GO-2026-6860 (golang.org/x/net, quadratic text-node coalescing in
html, fixed in v0.62.0) was reachable through dom.Parse (calls
html.Parse) and dom.ParseFragment (calls html.ParseFragment). x/net is
a DIRECT dependency here (no `// indirect` in go.mod), bumped to
v0.62.0.

GO-2026-6809 (crypto/x509 Certificate.Verify, fixed in go1.27.3) was
reachable from cmd/render because go.mod declared `go 1.27.2` - the
standard library is judged by the version the module claims. The
floor is named in three places and all three had to move: go.mod,
ci.yml's GO_VERSION, and the README's Go-version badge; raising go.mod
alone would leave CI building the vulnerable toolchain.

This repository had no vulnerability lane at all, so neither advisory
could ever have been reported - added .github/workflows/vuln.yml
(govulncheck judged by vulngate, on push/pull_request/a daily
cron/workflow_dispatch), mirroring go-webengine/browserproxy#26's
worked example.

Also found independently while verifying the sweep's report: the
nested bench/ module (chromedp-based, kept out of the main module's
CGO=0/6-arch build via its own `replace`) had drifted stale against
the parent - `go mod tidy` there failed outright, meaning half of this
repository's dependency graph was unreadable to any scanner. Renovate
updates the parent module and has no reason to touch a nested one it
does not itself declare. Fixed with `go mod tidy` in bench/, and added
a `bench` job to ci.yml (`go vet ./...`, not `go test` - bench's own
benchmarks are a main and running them measures the shared runner, not
this repository's code; no `-mod=mod`, which would hide the very
staleness the job exists to catch).

Co-authored-by: Claude Sonnet 5 <[email protected]>
tannevaled added a commit that referenced this pull request Nov 7, 2026
#27)

* security: klauspost/compress v1.20.2 and brotli v1.2.7, both REACHABLE

Found by the nightly lane this repository gained in #26, one night after
the advisories were published. It went red on 2026-11-06 and again on
11-07; the report it keeps on failure named both.

  GO-2026-6888  github.com/andybalholm/brotli v1.2.6 -> v1.2.7
                CVE-2026-111497. "A small crafted brotli stream containing
                many empty meta-blocks with new Huffman trees can trigger
                hundreds of megabytes of allocations per kilobyte of input,
                leading to denial of service."
    session.go:106:29  engineRender -> engine.Engine.RenderWithLinks
                       -> brotli.Reader.Read
    server.go:155:62   dispatch -> brotli.decodeError.Error

  GO-2026-6865  github.com/klauspost/compress v1.18.7 -> v1.20.2
                CVE-2026-116591. Panic on two-block frames in fse/huff0.
    server.go:115:22   Server.HandlerListener -> grpc.NewServer
                       -> fse.Decompress
    server.go:115:22   Server.HandlerListener -> grpc.NewServer
                       -> huff0.ReadTable

⛔ THE FIRST TRACE IS THE RENDER PATH. brotli.Reader.Read is reached from
engineRender, which is this program decoding a page it fetched: the exact
symbol the advisory names, on the job the proxy exists to do, against
content nobody here controls. The second pair is the gRPC server's
compression. Both are denial of service on untrusted input, and both have
released fixes.

⛔ AND BOTH WERE `// indirect`. Renovate has no version string to compare
or rewrite for a requirement the module does not declare for itself, so
nothing was watching either of them -- the third time this mechanism has
hidden a reachable advisory here, after x/net in #26.

Measured, same tree, same govulncheck, GOOS=linux GOARCH=amd64:

  before   2 vulnerabilities, 4 symbol traces (above)
  after    No vulnerabilities found. 0 affected.

The lane's own gate run by hand: govulncheck wrote 889 KB of JSON and
vulngate exited 0, judging the one remaining finding a notice (required,
not called, no RELEASED fix -- the upstream fix is a pseudo-version).

gofmt clean, go build ./... ok, go vet ./... silent, go test ./... 2
packages ok.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

* vuln.yml: two dashes for vulngate, before the shim goes away

Running the gate by hand to check the fix above, vulngate said:

  ⚠ vulngate: -go now takes two dashes; read as --go. The standard flag
    package accepted one dash and pflag does not. Update the call site:
    this shim will not outlive the next major.

vulngate moved to cobra. The single dash still works, through a shim that
announces its own removal on every run — and ⛔ a warning inside a GREEN
lane is a warning nobody reads. It would have become a red night on a
major bump, for a reason unrelated to anything this lane measures.

Verified rather than assumed, 2026-11-07 against vulngate@latest: `--go`
is accepted and silent, same verdict, same exit status.

The call site pins nothing: this lane installs vulngate@latest on purpose
(a pin inside a `run:` block is read by no dependency manager and rots —
go-gitsafe/gitsafe carried one 36 minor versions behind), so the shim's
removal would have arrived here without a commit of ours.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

---------

Co-authored-by: Claude Opus 5.5 <[email protected]>
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