browser-context-infer

Purpose

A sink's severity is fixed, but its viability depends on the page it runs on. el.innerHTML = userData is RCE on a no-CSP page, dangling-markup only on script-src 'none', and inert under enforced Trusted Types. This skill assembles the page-context evidence we have and writes targets/<name>/browser_context.json for downstream consumers.

Plan: plans/ARCHITECTURE_EVOLUTION.md §4.

When to run

Run once per target, after at least one of:

• caido-capture (real response headers captured),
• js-harvest (raw HTML present),
• passive-listen (runtime headers captured in runtime/*/state.json),
• js-index (so Trusted Types policy tags exist in the per-target DB).

Re-run when new evidence lands (e.g. you captured a fresh Content-Security-Policy header you didn't have before).

Inputs scanned (all optional)

Source	What it contributes
targets/<name>/raw/_response_headers.txt / .json	Real CSP headers
targets/<name>/raw/*.html (top-level only)	<meta http-equiv="CSP">
targets/<name>/runtime/*/state.json	passive-listen evidence — headers + policy bodies
targets/<name>/index/frameworks.json	Framework detection (used by rendering inference)
targets/<name>/db/js_analyzer.db	trusted_types_create_policy tags

Missing inputs degrade gracefully — every viability factor defaults to 1.0 (no change to chain scoring).

Steps

1. Run the driver:

   python3 bin/infer_browser_context.py <target_name>
   

   Output: targets/<name>/browser_context.json.

2. (Optional) Supply a CSP you captured manually:

   python3 bin/infer_browser_context.py <target_name> \
       --explicit-csp "default-src 'self'; script-src 'self' 'nonce-XXX'; require-trusted-types-for 'script'"
   

3. Enable viability scoring for the next chain extraction:

   export JS_ENABLE_SINK_VIABILITY=1
   export JS_ENABLE_SANITIZER_REALITY=1        # optional — Sprint A §5 sanitizer reality
   

   The reporter (tlx/modules/js_analyzer/reporter.py) multiplies the chain score by the viability factor and annotates demoted chains with viability_breakdown. Chains scoring below JS_VIABILITY_DROP_THRESHOLD (default 0.10) drop out of the hot list.

4. Sanity-check the output:

   jq '.csp | {source, raw, unsafe_inline_allowed, unsafe_eval_allowed, trusted_types_required}' \
       targets/<name>/browser_context.json
   jq '.trusted_types' targets/<name>/browser_context.json
   jq '.rendering' targets/<name>/browser_context.json
   

What it produces

targets/<name>/browser_context.json:

{
  "csp": {
    "raw": "default-src 'self'; ...",
    "source": "header",
    "script_src": ["'self'", "'nonce-XXX'"],
    "trusted_types_required": true,
    "unsafe_inline_allowed": false,
    "unsafe_eval_allowed": false,
    "confidence": 1.0
  },
  "trusted_types": {
    "enforced": true,
    "policies": ["default", "dompurify"],
    "policy_evidence": [{"qname": "...", "file": "...", "line": 42}]
  },
  "rendering": {
    "model": "ssr_then_csr",
    "framework": "nextjs",
    "hydration": true
  },
  "sandbox_iframes": [],
  "evidence_files": ["raw/_response_headers.txt", "index/frameworks.json"]
}

Boundary conditions

• No CSP evidence anywhere: viability falls back to 1.0 for every sink — no impact on scoring.
• TT enforced + TT-guarded sink (innerHTML etc.): viability factor drops to ~0.05.
• script-src 'none' without unsafe-eval: eval-family sinks drop to ~0.05; html-parser sinks stay at ~0.20 (CSS injection / dangling-markup still possible).
• Framework override: dangerouslySetInnerHTML under React keeps 1.0 (explicit opt-out); angular_inner_html_binding without bypass drops to 0.4 (Angular sanitizes by default).

Idempotency

Safe to re-run. The output file is fully regenerated each time. No state mutation outside targets/<name>/. No network. No LLM. No DB writes other than reads.

Failure modes

• Missing per-target DB: Trusted Types policy info is skipped, everything else proceeds.
• Malformed JSON in runtime/*/state.json: that file is skipped, others continue.
• Empty raw/: only meta-CSP is unavailable; header inputs still work.