Summary

Coraza extracts a multipart file upload's filename using Go's standard-library mime.ParseMediaType, which only decodes the RFC 5987/6266 extended filename* Content-Disposition parameter when its declared charset is exactly us-ascii or utf-8. Any other charset — including iso-8859-1, which RFC 5987 explicitly permits — is silently dropped by the stdlib, with no error and no fallback signal. This lets an attacker present one filename to Coraza and a different one to the backend application in the same request.

Affected variables

FILES (documented as "the original filenames as submitted by the client in the multipart upload"), FILES_SIZES, and the file-vs-field routing decision that determines whether a part is tracked as a file at all.

MULTIPART_FILENAME is not affected. It is declared in Coraza but is not populated by any code path in the affected versions.

Details

internal/bodyprocessors/multipart.go's originFileName calls mime.ParseMediaType on the raw Content-Disposition header and reads dispositionParams["filename"], which then populates FILES/FILES_SIZES for parts recognized as file uploads. Per RFC 7578 §4.2 and RFC 6266, when both filename and filename* are present, filename* is authoritative. Go's mime.ParseMediaType implements this precedence internally (decode2231Enc in mediatype.go), but only for filename* values whose charset token is us-ascii or utf-8:

// mime/mediatype.go
if charset != "us-ascii" && charset != "utf-8" {
    // TODO: unsupported encoding
    return "", false
}

When decode2231Enc returns false, the stdlib silently leaves whatever was parsed for the plain filename parameter untouched (or leaves filename absent entirely if only filename* was supplied) — there is no error, no partial-parse indicator, nothing a caller can detect. The stdlib also discards the RFC 5987 language component of filename* entirely; it is not recoverable from the parsed params map at all.

Verified behavior (prior to the fix)

In the second row, any rule inspecting FILES for a dangerous extension only ever sees safe.jpg, while a backend that follows RFC 7578 precedence and accepts iso-8859-1 (an RFC-5987-legal, non-exotic charset) uses shell.php as the filename. In the third row, the upload skips file-specific rule scope and size-limit tracking (FILES_COMBINED_SIZE) entirely.

Impact

A rule set that blocks file uploads by extension or name via FILES (or scopes rules to FILES/FILES_NAMES) can be bypassed by supplying the real filename via filename* under any charset other than utf-8/us-ascii — iso-8859-1 is sufficient and is a standards-compliant choice, not an obscure or malformed one — optionally alongside an innocuous decoy filename. A part carrying only filename* additionally escapes file-vs-field routing, so FILES, FILES_NAMES and FILES_COMBINED_SIZE never see it.

The bypass is deterministic against Coraza and requires no authentication or user interaction. Its effect is bounded to filename-derived inspection: request body content, ARGS and headers are still inspected, and a part carrying only filename* is routed into ARGS_POST. Coraza itself neither stores nor serves the uploaded file, so whether an evaded rule was preventing a change of state depends on the backend's own filename* precedence handling and upload handling. This is scored as a scope-changing bypass with low integrity impact (S:C, I:L) rather than a direct compromise.

Proof of Concept

Content-Disposition: form-data; name="upload"; filename="safe.jpg"; filename*=iso-8859-1''shell.php

Processed through internal/bodyprocessors/multipart.go prior to the fix, FILES for this part contained safe.jpg. A SecRule FILES "\.php$" (or similar extension-blocklist rule) did not fire, while a backend honoring RFC 5987/7578 filename* precedence uses shell.php as the filename.

Resolution

Fixed by locating and decoding filename* independently of mime.ParseMediaType's charset restriction (a quoted-string-aware scanner over the raw header). The fix also populates MULTIPART_FILENAME/MULTIPART_NAME, which were previously declared but unpopulated.

Both readings are exposed, not just filename*. An earlier version of this fix made filename* unconditionally authoritative over the plain filename, per RFC 7578 §4.2 — this closed the PoC above but relocated the same bypass: swap which field carries the real name (filename="shell.php"; filename*=iso-8859-1''safe.jpg) and a backend that resolves the plain filename instead is missed, since the discarded reading never reached FILES at all. Fixed in 89399e67: when a well-formed filename* picks a different value than the plain filename, both are added to FILES/FILES_SIZES/MULTIPART_FILENAME for the same part (one temp file, one byte count, two names). A SecRule FILES "\.php$" now catches the payload regardless of which field carries the real name.

Coraza does not select a "safe" charset allowlist on the operator's behalf. Instead, filename*'s declared charset and language are exposed as their own rule-matchable collections, MULTIPART_FILENAME_CHARSET and MULTIPART_FILENAME_LANGUAGE, alongside MULTIPART_FILENAME, so operators can enforce their own allowlist. Design credit: @airween.

For Content-Disposition: form-data; name="upload"; filename="safe.jpg"; filename*=UTF-8''shell.php:

MULTIPART_FILENAME:upload          = [shell.php, safe.jpg]   (filename* first, plain filename kept alongside)
MULTIPART_FILENAME_CHARSET:upload  = UTF-8
MULTIPART_FILENAME_LANGUAGE:upload = ""           (empty string; language is optional)

Rule writers can enforce a charset allowlist directly. Use an anchored regex
rather than @within:

SecRule MULTIPART_FILENAME_CHARSET:upfile "!@rx ^(?:utf-8|iso-8859-1|us-ascii)$" \
    "id:199,phase:2,t:none,t:lowercase,deny"

!@within utf-8,iso-8859-1,us-ascii looks equivalent and is not. @within
treats its parameter as the haystack and the variable as the needle, so an
empty value is trivially contained in it and the negation never fires. A
filename* that declares no charset at all — filename*=''shell.php, which
RFC 5987's grammar does not permit but which Coraza accepts and exposes rather
than rejecting — therefore passes an @within allowlist untouched. Measured:

Both spellings stay quiet on a part with no filename* at all, because the
collections are then absent rather than empty and the rule never evaluates.
That is what makes an allowlist rule safe to apply to all traffic rather than
only to known upload fields.

Discrepancies are no longer dropped silently

The same silent-fallback pattern this advisory describes existed on two neighbouring paths: a part whose Content-Disposition was present but unparseable, and a part carrying a filename* that did not match the charset'[language]'value shape at all. Both were reduced to a part with no filename — routed into ARGS_POST, outside FILES-scoped rule coverage, with nothing to signal that a filename had been discarded.

Both now raise MULTIPART_STRICT_ERROR (rule 200003 in CRS). A part with no Content-Disposition at all is not flagged: it simply carries no filename. mime.ParseMediaType accepts every real-world filename shape tested — raw UTF-8 and emoji, spaces, Windows backslashes, a bare %, unquoted values — so this signal does not reach legitimate uploads.

New variable: MULTIPART_DUPLICATE_PART_HEADER

Set to 1 when a part repeats a header (two Content-Disposition headers) or repeats a parameter inside its Content-Disposition (two filename parameters). A duplicate is precisely where backends disagree about which value wins, so the duplicate itself is the signal; it also contributes to MULTIPART_STRICT_ERROR.

SecRule MULTIPART_STRICT_ERROR "@eq 1" "id:201,phase:2,deny,t:none,chain"
  SecRule MULTIPART_DUPLICATE_PART_HEADER "@eq 1"

ModSecurity adds the same variable in its fix for GHSA-5pww-8rfg-9crf and lists it in the recommended audit-log format as DH. Coraza implements it under the same name so that a rule set referencing it parses on both engines.

Note on percent-decoding

Coraza percent-decodes the filename* value, so MULTIPART_FILENAME and FILES hold the name the application will actually resolve. This matters for rules that match on file extension: CRS 933110 and 944140 target FILES with t:none,t:lowercase,t:removeWhitespace and perform no URL decoding, so an encoded separator such as filename*=UTF-8''shell%2Ephp would evade them if the raw value were stored instead. ModSecurity's fix stored the raw, still-encoded value when this was first written; after the difference was raised on its advisory it now percent-decodes as well, so the two engines agree on this point.

Three further discrepancies found in post-fix review

Reviewing the fix (23446b5e) against test coverage turned up three more cases where Coraza's parsed value could disagree with what a backend resolves, closed in f30df398:

• An unresolvable % escape in filename* (filename*=UTF-8''safe.jpg%ZZ) is still kept as-is rather than substituting a decoy plain filename, but now also raises MULTIPART_STRICT_ERROR. Without this, the discrepancy was silent whenever a backend decoded the same escape differently or rejected it outright.
• A quoted filename* value (filename*="UTF-8''shell.php") is unwrapped before parsing. RFC 5987 does not permit filename* to be a quoted-string, but a general Content-Disposition parser such as Go's mime.ParseMediaType accepts a quoted value for any parameter. Without unwrapping, the literal quote characters ended up in MULTIPART_FILENAME/MULTIPART_FILENAME_CHARSET, so an anchored rule (e.g. \.php$) did not match a filename the backend resolved cleanly.
• MULTIPART_FILENAME, MULTIPART_FILENAME_CHARSET, MULTIPART_FILENAME_LANGUAGE and MULTIPART_NAME are multi-valued collections: two parts sharing one name (e.g. a multi-file upload field) each keep their own filename instead of the later part overwriting the earlier one. ModSecurity's fix for GHSA-5pww-8rfg-9crf independently identifies and fixes the same class of bug, describing it as "a second, distinct bypass," by keeping the equivalent variables multi-valued.

The precedence decision itself was found to relocate the bypass

Reviewing filename*-as-authoritative, @jptosso demonstrated (with a repro against Go's own mime/multipart as the reference backend) that this precedence choice doesn't close the two-sided decoy: it only decides which of the two payload shapes above is caught, not both. Checked against ModSecurity v3's actual fix for GHSA-5pww-8rfg-9crf: it makes the identical unconditional-precedence choice (single value, filename* wins, plain filename discarded), with no record of the two-sided case being considered there either — this is not a case of Coraza deviating from a more careful upstream fix.

Fixed in 89399e67 by keeping both readings rather than picking one (see Resolution above). Test coverage for the original PoC direction only ever exercised the filename="safe.jpg"; filename*=...''shell.php order; the swapped order (filename="shell.php"; filename*=...''safe.jpg) is now a dedicated regression case, since that gap in coverage is exactly why the one-sided version shipped in the first place.

Patched in 3.8.1

The 3.8.0 fix was incomplete. 3.8.1 completes it: an RFC 2231 numbered continuation (filename*0=, filename*0*=, ...) next to a bare filename* parameter now sets MULTIPART_STRICT_ERROR, and the Content-Disposition duplicate-parameter check added in 3.8.0 is now linear instead of quadratic in the number of parameters. Upgrade to 3.8.1; 3.8.0 is listed as affected.

Severity (revised 2026-10-02)

CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:C/C:N/I:L/A:N (4.0, Medium).

Attack Complexity is High because the bypass depends on a specific backend behaviour outside the attacker's control. Go's mime.ParseMediaType and PHP resolve the same filename Coraza did, so they are not affected. The discrepancy exists only for backends that decode a non-UTF-8 filename* or treat a filename*-only part as a file, such as Werkzeug/Flask and busboy-based Node.js backends (Express with multer, Fastify). The continuation case fixed in 3.8.1 is resolved only by Werkzeug. The previous vector scored Attack Complexity Low (5.8).

Impact metrics follow the convention used across Coraza's WAF-bypass advisories: the vulnerable component is Coraza, but the impact lands on the protected application, so Scope is Changed. The bypass hides a payload from inspection; the application still has to be vulnerable to it, so Integrity is Low and Confidentiality is not scored separately.

AI involvement in this section: Claude Opus 5.5 (Anthropic), via Claude Code, re-derived the CVSS vector from the project's triage guidance (AGENTS.md, "CVSS preconditions get verified, not copied from the report") and drafted this text. A human maintainer (fzipi) chose the S:C/I:L impact convention and directed this update.