Security disclosure — Clubhouse room reactions render unvalidated text
To: Clubhouse security team From: [YOUR NAME / HANDLE — third-party Clubhouse client developer] Date: [DATE] Severity (our assessment): Medium — abuse / social-engineering vector, no account compromise observed
Summary
The room reaction feature accepts an arbitrary string in the reaction's
emoji field and renders it to every participant without validating that the
value is actually an emoji. A crafted reaction causes long, arbitrary text to
burst across the screens of everyone in the room, using the normal reaction
animation. We have observed this in the wild: a reaction carrying placeholder
"leaked data" text ([DEMO] Leaked … IBAN:… CVV:… SSN:…) sprayed as white text
over a live room, which participants reasonably mistook for an account breach.
No data is actually exposed by this — the text is attacker-supplied and fake — but it is an effective scare / impersonation / spam vector because it looks like a system-level leak and originates from a real participant's tile.
Root cause
Two layers each miss a check:
-
Server — reaction send endpoint. The reaction submission (
POST /emoji_reaction,{channel, emoji}) does not appear to validate thatemojiis a member of the room's advertisedemoji_reaction_options, nor that it is an emoji at all. Any string inemojiis accepted and fanned out to the room's realtime feed as anew_channel_reactionevent. -
Client — reaction renderer. The official client draws whatever string arrives in the reaction event as if it were an emoji glyph, with no length cap and no emoji check. A text string therefore renders as text, repeated and rotated by the burst animation, anchored on the sender's tile.
The advertised emoji_reaction_options list is enforced only by the sending
client's own picker UI, so a request made outside that UI bypasses it entirely.
Impact
- Scare / social engineering: fake "leaked credentials" text that looks like a breach, as observed.
- Impersonation / harassment / spam: arbitrary text attributed to the sender's avatar, shown to all participants, repeatable.
- Client stress: because the client caps neither length nor (as far as we can tell) rate, large or rapid payloads may degrade the room UI on participants' devices.
We did not observe any account takeover, data exfiltration, or privilege escalation. The "leaked" content is fabricated by the attacker.
Reproduction (mechanism, not a weaponized payload)
We are deliberately not including a ready-to-run exploit request. The mechanism:
- A reaction event of the normal shape reaches the room feed:
{ action: "new_channel_reaction", action_user_profile: {…}, channel, message_id, <content> }. - The
<content>carries a non-emoji string instead of one of the room'semoji_reaction_options. - Every participant's client renders that string via the reaction burst.
We can reproduce the client rendering half entirely offline, with no network and no live room, by feeding a crafted event into a reaction parser — confirming that the string is drawn verbatim. We are happy to share that in-process test privately on request.
Suggested remediation
- Server: reject a reaction whose
emojiis not in that room'semoji_reaction_options(or, at minimum, is not a short emoji-only string). This is the single highest-value fix, since it stops the fan-out at the source. - Client: treat the reaction field as untrusted — cap its length, require it to be an emoji before rendering, and rate-limit bursts per sender.
- Defense in depth: apply the same validation on any other field a client renders from a realtime event.
What we changed on our side
Our client is third-party and read-mostly, but we hardened both directions:
We validate that any reaction we send is a genuine emoji.Since 2026-10-17 our client deliberately sends custom text reactions too. Each one is limited to a single line of at most 64 characters, with control, bidi-override and line-separator characters refused. That rules out the multi-line, layout-spoofing payloads, but not short text.- We cap and escape incoming reaction text before rendering, so a hostile payload cannot grow our UI unbounded or be interpreted as markup.
Contact
Happy to provide the offline reproduction test, wire captures of the observed event, or any other detail that helps. Please let us know your preferred disclosure channel and timeline.