All postsThe postMessage listener that trusted everyone
No origin check, innerHTML sink, public page. Accepted as Medium. I argued for High and lost.
XSS · 3 min read
An HR self-service portal. Almost every route needs a login, but the home page
returns 200 to anyone, and the portal template injects this on every page:
var eventMethod = window.addEventListener ? "addEventListener" : "attachEvent";
var eventer = window[eventMethod];
var messageEvent = eventMethod === "attachEvent" ? "onmessage" : "message";
eventer(messageEvent, (e) => {
if (e.data && e.data.text)
document.getElementById('widget-title-label').innerHTML = e.data.text;
});
No e.origin check, and innerHTML where textContent would do. Either one is a
bug on its own. Together it's unauthenticated script execution in an
authenticated session.
The split is what makes it work: I reach the sink with no account because the
page is public, and the victim brings the cookies. Neither side needs what the
other has. The script lands in the portal origin, so it reads any same-origin
route that session can reach.
Delivery#
const w = window.open('https://portal.example.com/');
setTimeout(() => {
w.postMessage({ text: '<img src=x onerror="alert(origin)">' }, '*');
}, 3000);
What actually costs time here:
- Popup, not iframe. An iframe is quieter and needs no gesture, but dies if
the target sets
X-Frame-Options or frame-ancestors. Check first, then fall
back to window.open, which needs a real click or the blocker eats it.
- The load race.
window.open returns immediately; the listener doesn't exist
until their scripts run. A fixed setTimeout is fine for a proof and flaky in a
live demo. Usual reason a correct payload "doesn't work" in front of triage.
innerHTML won't run <script>. It parses and does nothing. Use a tag that
carries its own handler.
targetOrigin: '*' on the sender isn't the bug. The missing e.origin check
on the receiver is. Worth noting separately if the message carries anything
sensitive.
Prove the sink with alert(origin) before building anything on top of it. If
that doesn't fire, nothing downstream matters.
Finding more of them#
addEventListener\(['"]message['"]
onmessage\s*=
\.origin
Read each listener and ask two things in order: is e.origin checked, and where
does e.data end up.
| Sink |
What you get |
innerHTML, outerHTML, insertAdjacentHTML |
HTML injection, script via an event handler |
document.write |
Same, and it can clobber the page |
eval, Function, setTimeout(string) |
Direct execution |
location, location.href |
Redirect, javascript: if unfiltered |
localStorage / sessionStorage |
Stored payload that fires on a later page |
getEventListeners(window) in Chrome lists them without reading a bundle. Weak
checks count too: e.origin.indexOf('portal.example.com') > -1 matches
portal.example.com.attacker.net. Compare the whole string.
Template-injected bugs are worth enumerating across every site built from that
template. This handler sat on prod, a UAT box, and an unrelated travel portal. It
also tells the vendor to fix the template rather than the page.
Severity#
I scored it 8.2: AV:N/AC:L/PR:N/UI:R/S:C/C:H/I:L/A:N. Accepted as 5.4 Medium.
I filed a reconsideration. The XSS could read a live chat widget's session
identifier out of the page, and that identifier re-minted a valid chat token with
no authentication, so the label overwrite was really a foothold into an employee's
support conversation. Reasonable argument.
It didn't move the score, and the reason is in my own text:
My reconsideration, verbatim
"Happy to provide the full request/response pairs, the idempotency proof across
multiple days, and the transcript-read PoC on request."
I described the chain and offered the evidence instead of attaching it. Nobody
re-scoring a closed report is going to open a correspondence to collect proof I
already had on disk. Attach it or don't bother.
Fix#
const ALLOWED = ['https://portal.example.com'];
eventer(messageEvent, (e) => {
if (!ALLOWED.includes(e.origin)) return; // exact match, not indexOf/startsWith
if (e.data && e.data.text)
document.getElementById('widget-title-label').textContent = e.data.text;
});
Reported under a VDP and accepted. Target isn't named and the identifiers are
genericised: "accepted" isn't permission to publish.