CVE-2026-104252: Vitess vttablet /twopcz reflected XSS in vitess.io/vitess

From: xylove21 To: security@vitess.io Subject: [Security Report] Vitess vttablet /twopcz reflected XSS Date: Tue, 24 Jun 2026 14:30:00 +0800

Summary

The twopczHandler in go/vt/vttablet/tabletserver/twopcz.go (line 185) reflects the Action and dtid form values directly into the HTML response body via fmt.Fprintln(w, msg) — without HTML escaping. The response is served as text/html; charset=utf-8, and Vitess's default acl policy is allow-all, so the endpoint is reachable by anyone who can hit the vttablet's HTTP port (default 15999, multiplexed with gRPC and MySQL wire-protocol ports).

A crafted Action or dtid value yields a reflected XSS that executes in the operator's browser with same-origin access to the vttablet's debug endpoints. This includes /queryz (live query plans), /querylogz (recent queries with bound values), /txlogz (transactions), /debug/env (environment variables — may include MySQL/topology credentials), /debug/vars (expvar stats), and /debug/pprof/ if enabled.

The bug was introduced before v20.0.0 (long-standing) and was silently fixed in commit 3140c1846c (merged 2026-06-09, PR #20255, "tabletserver: escape reflected form values in twopcz handler"). The fix replaces fmt.Fprintln(w, msg) with fmt.Fprintln(w, html.EscapeString(msg)). No released version of Vitess contains the fix — the fix lives only on main (current tip 33a3c0fafe69, 2026-06-23). There is no CVE, GHSA, or OSV entry for this specific bug (verified 2026-06-24 against GitHub Advisory Database, OSV.dev, and NVD).

Impact

| Field | Value | |-------|-------| | Type | Reflected Cross-Site Scripting (XSS) | | CVSS 3.1 | AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:N/A:N = 7.7 HIGH | | Attack Vector | Network (vttablet HTTP port 15999, multiplexed with gRPC/MySQL) | | Attack Complexity | Low (craft URL, send to operator) | | Privileges Required | None (default acl policy is allow-all) | | User Interaction | Required (operator must visit crafted URL) | | Scope | Changed (XSS chain reaction to other vttablet debug endpoints) | | Confidentiality | High (query text, env vars, expvar stats) | | Integrity | None (XSS itself doesn't change state) | | Availability | None (XSS itself doesn't disrupt service) |

Affected versions (verified)

All released versions are vulnerable. The fix lives only on main:

| Tag | Released | twopcz.go line 185 | Status | |---|---|---|---| | v20.0.0 | 2024-06-27 | fmt.Fprintln(w, msg) | BUGGY | | v20.0.1 - v20.0.6 | 2024 | fmt.Fprintln(w, msg) | BUGGY | | v22.0.4 | 2026-02-27 | fmt.Fprintln(w, msg) | BUGGY | | v23.0.3, v23.0.4 | 2026-02 / 2026-05 | fmt.Fprintln(w, msg) | BUGGY | | v24.0.0, v24.0.1 | 2026-04 / 2026-05 | fmt.Fprintln(w, msg) | BUGGY | | main (33a3c0fafe69) | 2026-06-23 | fmt.Fprintln(w, html.EscapeString(msg)) | FIXED |

Sensitive data exposed via the same-origin XSS chain

A successful XSS lands in the vttablet's origin. From there, the attacker can fetch (via fetch() from injected JavaScript) the following same-origin endpoints, all of which are unauthenticated by default:

- /twopcz — in-flight 2PC transactions (query text, parameters, dtid) - /queryz — live query plans - /querylogz — recent queries with bound values - /txlogz — transactions - /debug/env — environment variables (may include MySQL creds, topology credentials, encryption keys) - /debug/vars — expvar stats - /debug/pprof/ — CPU/memory profiles (if enabled)

Reproduction (independent, runnable PoC)

The PoC is a self-contained Go HTTP server that replicates the buggy twopczHandler from Vitess v20.0.0 verbatim (using the same line-185 code). It runs on 127.0.0.1:15080 and returns the same response shape as the real Vitess endpoint.

bash 1) Build the PoC (Go 1.23+) cd /path/to/finding/CVE-pending-twopcz-xss go build -o /tmp/vitess-twopcz-poc ./poc.go

2) Start the buggy server ADDR=127.0.0.1:15080 /tmp/vitess-twopcz-poc &

3) Reproduce — Test 1: bare GET (no params) curl -sS -i "http://127.0.0.1:15080/twopcz" → Content-Type: text/html; charset=utf-8, status 200

4) Test 2: XSS via Action param (URL-encode the script tag) curl -sS -i "http://127.0.0.1:15080/twopcz?Action=%3Cscript%3Ealert(1)%3C%2Fscript%3E&dtid=zz" → Response body contains literal "<script>alert(1)</script>(zz): completed." → Browser executes the script as same-origin code

5) Test 3: XSS via dtid param (attribute-context payload) curl -sS -i "http://127.0.0.1:15080/twopcz?Action=Demo&dtid=%3Cimg%20src%3Dx%20onerror%3Dalert(1)%3E" → Response body contains literal "<img src=x onerror=alert(1)>" → Browser fires the onerror handler as same-origin code

6) Run the 3x automated runner (writes evidence/run-{1,2,3}.log) bash ./poc/run.sh

PoC result excerpt (from evidence/run-1.log):

--- Test 2: XSS via Action=<script>alert(1)</script> --- HTTP/1.1 200 OK Content-Type: text/html; charset=utf-8 ... <h2>WARNING: Actions on this page can jeopardize data integrity.</h2> <script>alert(1)</script>(zz): completed.

PASS: payload reflected UNESCAPED (XSS exploitable in browser)

The PoC was run 3 times on the buggy server (all 3/3 PASS) and the same code is byte-identical to the v20.0.0 commit 8aeb274f6c8e line 185.

Root cause (technical detail)

go/vt/vttablet/tabletserver/twopcz.go (v20.0.0 through v24.0.1):

go // Line 185 fmt.Fprintln(w, msg)

msg is constructed earlier in the handler as:

go msg := fmt.Sprintf("%s(%s): completed.", action, dtid)

with action and dtid coming from the form values r.FormValue("Action") and r.FormValue("dtid"). The form values are URL-decoded (so a + becomes a space, %3C becomes <) but never HTML-escaped. The handler explicitly sets Content-Type: text/html; charset=utf-8 (via the <!DOCTYPE html> it writes first), so the browser executes any injected <script> or attribute-context payload.

The fix in main (fmt.Fprintln(w, html.EscapeString(msg))) closes the hole by HTML-escaping both action and dtid before reflection.

Suggested fix (already in main)

go // Recommended minimal patch (already in main, needs backport to all release branches): import "html"

func twopczHandler(w http.ResponseWriter, r http.Request) { // ... existing logic ... msg := fmt.Sprintf("%s(%s): completed.", action, dtid) fmt.Fprintln(w, html.EscapeString(msg)) // ← was: fmt.Fprintln(w, msg) }

Backport the fix to v20.0.x, v22.0.x, v23.0.x, and v24.0.x as patch releases. The fix is one line and a one-line import "html" addition, so the backport risk is minimal.

Workaround / mitigation (no patch available)

If you cannot upgrade to a fixed release immediately:

1. Restrict the vttablet HTTP port (default 15999) to localhost only. vttablet can run with --tablet-allow-stats=false and --tablet-grpc-port configured separately from the HTTP port, or you can firewall the HTTP port at the network layer. The debug endpoints are not needed by production traffic.

2. Set a non-default ACL policy. Vitess supports an acl policy that defaults to allow-all. Setting it to a JSON file that denies unauthenticated access to debug endpoints (see go/acl/acl.go::savePolicy) would block the XSS, but the default policy is currently "" (allow-all) and many users don't override it.

3. Use a reverse proxy (nginx, envoy) in front of the vttablet HTTP port with a strict allowlist of clients.

Cross-references

- The bug is not related to: - CVE-2024-53257 (/debug/querylogz HTML injection, fixed in v20.0.4 — different endpoint) - CVE-2026-27969 (backup MANIFEST path traversal) - CVE-2026-27965 / GHSA-8g8j-r87h-p36x (backup external decompressor RCE) - PR #20255 is the fix commit; PR #19486 (defense-in-depth) is a separate hardening change

Contact

- GitHub: @xylove21