Skip to content

go 1.27.4 for GO-2026-6859, and the CI pin reads go.mod - #28

Merged
tannevaled merged 1 commit into
mainfrom
go1274
Nov 15, 2026
Merged

tannevaled merged 1 commit into
mainfrom
go1274

Conversation

@tannevaled

@tannevaled tannevaled commented Nov 15, 2026 •

Copy link
Copy Markdown
Contributor

Two halves, both required: the published floor, and the CI pin that would otherwise refuse it.

1. go 1.27.3 → go 1.27.4

GO-2026-6859, published 2026-11-11, CVE-2026-97054, "HTTP/1 server crash due to MaxBytesReader header race in net/http", standard library, fixed in 1.26.11 and 1.27.4 — read at vuln.go.dev rather than quoted from notes.

The go directive is the only one of the three Go pins a consumer reads: the CI pin builds our artefacts, the toolchain stamped in a binary describes one executable, and this line is what go install honours under GOTOOLCHAIN=local. modfloor --floor 1.27.4 read the directive this module published at v0.2.1 as go 1.27.3.

⛔ And this one is a proxy — it serves HTTP as well as fetching it, so an advisory about an HTTP/1 server is this program's own class rather than a contract detail. Whether its handlers reach the racing path stays govulncheck's question, not this line's.

Why now: this repository is named in go-fleettools/fleettools/.github/vulnscan-exposed.txt, the nine with a network surface — a list that file says is cut by exposure rather than by convenience, because both repositories that actually reached an advisory in the fleet's 58 decoders were network-facing, and this was one of the two.

2. ⛔⛔ The workflow asked for this in writing, and nothing enforced it

env.GO_VERSION carried the literal "1.27.3" under this comment:

KEEP THIS EQUAL TO go.mod's go DIRECTIVE. […] Raising go.mod alone would leave the lane building the vulnerable one.

That was right, and it was a rule with nothing to enforce it. Two weeks later the raise arrives and the pin still says 1.27.3. The failure is not even a quiet older build: setup-go sets GOTOOLCHAIN=local, so the lane refuses with go.mod requires go >= 1.27.4. Measured the same day on five of six go-filesystems repositories — every job red at once, for a one-line change.

So the three steps now read go-version-file: go.mod and GO_VERSION is gone. The equality the comment requested is structural instead of requested, and the next raise needs no edit here. vuln.yml already read the file, which is where the shape came from.

⚠ Why nothing flagged it: gofloor reports a hard-coded go-version as a finding and classifies ${{ env.GO_VERSION }} merely as "expression". A literal one indirection away was invisible to the auditor written to find exactly this drift. Filed separately rather than fixed here.

Measured

actionlint clean, gofmt -l . silent, go build ./... and go test ./... pass, and no go-version: literal remains in any workflow of this repository.

Release level: MINOR

A raised floor genuinely breaks a consumer pinned below it with GOTOOLCHAIN=local: their build stops rather than quietly producing a vulnerable binary. The right failure, and still a break. The release is the fix, not the merge — a consumer reads the published go.mod.

🤖 Generated with Claude Code

… env var

GO-2026-6859, published 2026-11-11, CVE-2026-97054: "HTTP/1 server crash due to
MaxBytesReader header race in net/http", standard library, fixed in 1.26.11 and
1.27.4. Read at https://vuln.go.dev/ID/GO-2026-6859.json rather than quoted
from notes.

The `go` directive is the only one of the three Go pins a CONSUMER reads: the CI
pin builds our artefacts, the toolchain stamped in a binary says what one
installed executable was built with, and this line is what `go install` honours
under GOTOOLCHAIN=local. `modfloor --floor 1.27.4` read the directive this
module PUBLISHED at v0.2.1 as `go 1.27.3`.

⛔ AND THIS ONE IS A PROXY: it serves HTTP as well as fetching it, so an
advisory about an HTTP/1 SERVER is this program's own class rather than a
contract detail. Reachability remains govulncheck's question, not this line's.

WHY THIS REPOSITORY NOW: it is named in
go-fleettools/fleettools/.github/vulnscan-exposed.txt, the nine repositories
with a NETWORK SURFACE — a list that file says was cut BY EXPOSURE rather than
by convenience, because both repositories that actually REACHED an advisory in
the fleet's 58 decoders were network-facing, and this was one of the two.

⛔⛔ THE SECOND HALF, AND THE WORKFLOW ASKED FOR IT IN WRITING. `env.GO_VERSION`
carried the literal "1.27.3" with this comment above it:

    KEEP THIS EQUAL TO go.mod's `go` DIRECTIVE. [...] Raising go.mod alone
    would leave the lane building the vulnerable one.

That was right, and it was a rule with nothing to enforce it — two weeks later
the raise arrives and the pin still says 1.27.3. The failure is not a quiet
older build either: `setup-go` sets GOTOOLCHAIN=local, so the lane REFUSES with
`go.mod requires go >= 1.27.4`. Measured the same day on five of six
go-filesystems repositories, every job red at once.

So the three steps now read `go-version-file: go.mod` and GO_VERSION is gone.
The equality the comment requested is structural instead of requested, and the
next raise needs no edit here. `vuln.yml` already read the file, which is where
the shape came from.

⚠ WORTH KNOWING, because it is why nothing flagged this: `gofloor` reports a
hard-coded `go-version` as a finding, and classifies `${{ env.GO_VERSION }}`
merely as "expression". A literal one indirection away was invisible to the
auditor written to find exactly this drift. Filing separately.

actionlint clean, gofmt silent, go build and go test ./... pass, no
`go-version:` literal left in any workflow.

MINOR, not patch: a raised floor is a genuine break for a consumer pinned below
it with GOTOOLCHAIN=local — their build stops instead of quietly producing a
vulnerable binary.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@tannevaled
tannevaled merged commit 271be2f into main Nov 15, 2026
9 checks passed
@tannevaled
tannevaled deleted the go1274 branch November 15, 2026 15:26
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