CVE-2026-69246: Guzzle: Noncanonical host can bypass host-based checks

In affected versions, Guzzle gives a transport the request URI as text and supplies the Host header separately. The cURL handlers set CURLOPT_URL to the URI exactly as written and push that Host into CURLOPT_HTTPHEADER; StreamHandler does the same through fopen(). libcurl then parses the authority itself, percent-decoding it and, on an IDN-capable build, applying IDNA mapping, and uses the result to resolve, connect, name the TLS peer and address a proxy CONNECT, while the supplied Host suppresses the aligned one it would have generated. In http://127.0.0.%31/ the URI host is one filter_var() rejects as an IP literal, yet libcurl decodes it to 127.0.0.1 and reaches loopback with no DNS lookup while the server receives Host: 127.0.0.%31.

An attacker who influences a fetched URI can therefore reach a host the application’s checks excluded and read whatever it exposes of the response. The same divergence moves Guzzle’s own decisions onto a spelling the transport does not use: no_proxy selects proxy routing from the literal host, and RedirectMiddleware decides from it whether to strip Authorization and Cookie. The cookie middleware extracts Set-Cookie against the URI Guzzle produced, not the authority contacted, so for a raw divergent URI the cookie is stored under the URI host as written. Where Guzzle rewrote that URI but left a divergent Host, or where the caller supplied one, the cookie is stored under the canonical name and replayed by ordinary later requests to it. With a third-party UriInterface, a host of blocked.example.com@127.0.0.1 reaches 127.0.0.1 through all three handlers and generates Authorization: Basic from userinfo the application never wrote.

Exploitation requires the application to build a request URI from untrusted input and to make a host decision before handing it to Guzzle. Applications that only fetch URIs they construct themselves are not affected, and an exact allowlist of canonical names ordinarily fails closed; the exposure is to denylists, private-range and IP-literal checks, and any check that treats an unresolvable name as safe. The raw Unicode class needs an IDNA transformation somewhere: either a libcurl built with IDN support or Guzzle’s own idn_conversion, off by default on both branches, which rewrites the URI in Client::buildUri() before a handler sees it and leaves a prebuilt request’s explicit Host as written, while a request the client builds derives that header from the rewritten URI and produces no divergence. Noncanonical numeric spellings such as 127.1, 2130706433, 0x7f000001 and 0177.0.0.1 remain accepted after the patch and reach whatever the transport reads them as, loopback or a routable public host, and the cURL and stream handlers can differ, so a check comparing a host against an address as text stays bypassable. Guzzle does not offer SSRF protection, and neither cache poisoning nor cross-tenant compromise was established.