All postsThe display name was whatever I said it was
Three behaviours that were fine on their own. The argument that got it accepted wasn't about severity, it was about scope.
Auth · 3 min read
An employee HR portal embedded a hosted live-chat widget. The public,
unauthenticated page that loads the widget also publishes every identifier the
chat backend needs: org id, widget id, backend URL. That's not a leak, it's how
the widget bootstraps.
From those identifiers, anonymously:
GET /livechatconnector/v4/getchattoken/{orgId}/{widgetId}/{random-uuid} HTTP/2
Host: <tenant>.<chat-vendor>.com
HTTP/2 200 OK
{"ChatId":"19:acsV2_…@thread.v2","Token":"eyJ…","ExpiresIn":"86399", …}
A 24-hour token, and a live session now sitting in the company's support queue.
Then, straight against the vendor's chat API:
POST /chat/threads/{ChatId}/messages HTTP/2
Authorization: Bearer eyJ…
{"content":"…","type":"text","senderDisplayName":"HR Support Agent"}
201 Created. The agent's console renders that display name verbatim.
None of those three is a bug#
This is the part worth slowing down on, because it's where I nearly threw the
finding away.
Anonymous token issuance is by design. The whole point of a website chat
widget is that a visitor who has never logged in can start a conversation. An
endpoint that mints an anonymous session token without credentials is the feature.
The token working directly against the chat REST API is also by design. It's a
normal bearer token for a normal documented API. I'm not bypassing anything by
using it outside the widget.
senderDisplayName in the request body is the bug. It's supplied by the
caller and never reconciled against the identity in the token. The token's subject
is an anonymous user with no name attached, so there is nothing to contradict.
Whatever string I send is what the human agent sees, with no indicator in the UI
that it was caller-supplied rather than identity-derived.
So: an anonymous person can place messages attributed to "HR Support Agent", "IT
Security Team" or a named colleague into a live employee support conversation.
Check the boring explanation first
Two of those three behaviours are a product working correctly. If I'd written
the report around "the token endpoint needs no authentication" it would have
been closed as by-design in a sentence, and correctly. The finding only exists
because one field in the chain isn't validated.
The scope fight#
Triage pushed back immediately, and reasonably: that host belongs to the chat
vendor, not to us. Third-party infrastructure. Out of scope.
That objection is usually right, and it quietly kills a lot of real findings on
SaaS-heavy targets. I had two replies drafted in the folder — one arguing it was
in scope, one withdrawing the report. I came close to sending the withdrawal.
What worked was dropping the "your vendor has a bug" framing entirely. Three
moves:
- Show the in-scope asset loading it. Paste the actual
<script> tag from
the in-scope page, with the org and widget ids visible in its attributes. Not a
description of the integration — the tag.
- Show every input came from the in-scope asset. Each identifier in the repro
was read off that public page. Nothing was obtained from the vendor, guessed, or
carried in from another target.
- Locate the defect in the tenant's configuration, not the vendor's code.
Anonymous chat was switched on for this deployment, and the display name was
never constrained. Both are the customer's settings on their own instance.
That last one is the sentence that moves it. "Your vendor has a vulnerability" is
someone else's report. "Your configuration of your vendor exposes your employees"
is this program's report, and it is also just true.
- Read the bootstrapper
<script> tag. The ids you need are usually right there.
- Find the token endpoint and ask precisely what the token authorises, and for how
long.
- In every message or event body, find the fields describing who is speaking —
display name, avatar, role, sender id — and test whether any of them are
validated against the token.
- Then ask who renders it, and whether that human has any way to tell a supplied
name from a verified one.
Outcome#
Filed as High, chaining all three. Accepted as Medium, no bounty — it's a
disclosure programme.
One more thing, and it's the reason I went back to this at all. My own notes
recorded this report as withdrawn, out of scope, final disposition unverified,
and sat that way for over a month. The platform record said Accepted. I'd been
carrying a wrong number in my own ledger because I wrote down what I expected to
happen instead of checking what did.
The platform is the ledger. Your notes are a guess.