The Advisory Your Dependency Bot Cannot See — Seven CVEs, Zero GHSAs, and a Compiled-Graph Sweep
Table of Contents
I run my own dependency bot. You may have read that post — it was mostly a love letter to boring commits with receipts in the body. This is the darker companion piece: the month the boring pipeline caught something the famous pipeline couldn’t.
Seven advisories. All reachable in my compiled binary. All of them invisible to Dependabot — not because the bot was misconfigured, but because the data source it reads simply didn’t have them.
What I actually had
Here’s the commit message, verbatim in shape:
Bump x/crypto 0.55.0->0.57.0 + x/net 0.58.0->0.60.0 (security) (#137)
Clears 7 OSV/GO advisories in COMPILED (reachable) packages:
- x/crypto/ssh DoS pair: GO-2026-6354 (CVE-2026-78662, deadlocked
undecided channel), GO-2026-6355 (CVE-2026-56855, established channel)
— ssh is compiled via ollama/api auth chain (4 packages).
- x/net HTTP/2 server stack: GO-2026-6603 (Trailer-header memory
exhaustion), GO-2026-6610, GO-2026-6611, GO-2026-6612, GO-2026-6617
(HPACK encoder race crash). x/net/http2 compiled; mule serves HTTP.
Two families, one interesting each.
The SSH pair. Two denial-of-service bugs in golang.org/x/crypto/ssh: a deadlock on an undecided channel, and a problem on an established one. The uncomfortable part is how SSH got into my binary at all. Mule doesn’t speak SSH. It talks to Ollama. The ollama/api client’s auth chain drags in x/crypto/ssh — four packages deep — and once it’s compiled, its bugs are your bugs. This is why “I don’t use that package” is not a security posture. The linker is; reachability is.
The HTTP/2 quintet. Five advisories in the x/net HTTP/2 server stack: trailer-header memory exhaustion (CVE-2026-78659), three siblings in the same neighborhood, and an HPACK encoder race that crashes. I serve HTTP. That’s the whole argument.
The blind spot: Go-team-only advisories
The reason this is a post and not a shrug is where the advisories lived. All seven were Go-team-only advisories — GO- entries in the OSV database, no GitHub Security Advisory. I checked: lookups for those CVE IDs in GitHub’s advisory database return zero rows.
Dependabot reads GitHub’s advisory database. No GHSA, no alert — structurally, not configurally. There is no setting. No webhook to enable. My green Dependabot badge and my seven reachable CVEs coexisted peacefully for weeks. Every “dependabot says we’re clean” dashboard on earth is making the same trade: coverage of what GitHub knows about, which is a subset of what the Go security team knows about, which is a subset of reality.
govulncheck reads OSV, so the standard library and x/* Go-team advisories do land there — that’s how a compromised-intermediate x509 fix reached me a week earlier. But plenty of teams run Dependabot and call it done. If that’s your whole stack, GO--only advisories are a hole shaped exactly like your blind spot.
The fix: scan the compiled graph, not go.sum
The sweep that found these wasn’t clever. It asked OSV about every package in the compiled graph — what go build actually links — rather than every module go.mod mentions.
# The compiled graph, not the module graph:
go list -deps ./... | grep -v '^\(internal\|github.com/mule-ai\)' > pkgs.txt
# Ask OSV about each versioned package (batched, real shape):
curl -s -X POST https://api.osv.dev/v1/query -d '{
"queries": [{"package": {"name": "golang.org/x/net", "ecosystem": "Go"},
"version": "v0.58.0"}]
}' | jq '.results[].vulns[] | {id, aliases, summary}'
The distinction is the whole game. go.mod lists what you require; go list -deps is what you ship. go-git’s test-only dependencies don’t page me. x/net/http2 pages me because I called ListenAndServe. My pipeline runs this sweep independently of Dependabot on purpose: two scanners, two data sources, no shared blind spots.
The coupling trap
One wrinkle worth pinning to your memory: the fixes moved together whether I wanted them to or not. x/net v0.60.0 (the HTTP/2 fixes) requires x/crypto >= v0.57.0 — which happens to be the version carrying the SSH fixes (the fix threshold for the pair is 0.56.0). So one commit bumps both, plus transitives dragged by go mod tidy: x/sys 0.47→0.48, x/text 0.41→0.42.
Two lessons in that:
- Don’t cherry-pick “just the advisory I care about” in the
x/*ecosystem. The Go subrepos version-lockstep for a reason; fight it and you’ll build something nobody has ever run. - When security bumps come in a set, verify the set.
go build -mod=readonly ./...andgo test -mod=readonly ./...(15 packages, 0 failures, matching baseline), then re-run the OSV sweep over the compiled graph: reachablex/*hits, 7 → 0. The receipt goes in the commit body. The receipt is the point.
What to steal
- Add OSV (or
govulncheck, which shares its feed) to CI today if you only run Dependabot. The Go-team-only class is real and quiet. - Scan the compiled graph (
go list -deps ./...), notgo.mod. The module graph lies by omission in both directions. - Ask “why is this in my binary?” when a weird package shows up in deps. Mine was SSH-via-Ollama; yours might be a zip parser via an image library.
- Bump coupled packages as one commit, and let
go mod tidyhave the last word. - Write the before/after scan counts in the commit message. “7 → 0” is auditable; “updated deps” is a rumor.
Your dependency bot sees what its database contains. Mine happens to have two databases now. The advisories your bot cannot see are not the ones you should worry about — they’re the ones you’d never think to look for.