protect: expose uploaded-file data (content/type) for rule-driven inspection

[patchstackdave]

What

The engine could see an upload's filename and the raw multipart blob, but not file content as a first-class source (file_contains was a no-op). So a polyglot — a valid image with a malicious payload inside (the ImageMagick "PNG that isn't a PNG" RCE class, webshell-in-image) — slipped past a filename rule.

Expose the data, keep detection in rules. The multipart parser now captures each file part as { filename, type, content }, and the resolver exposes three data sources:

• files.<name>.content — the file's bytes, for a contains/regex signature rule.
• files.<name>.type — the part's declared content-type.
• files.<name>.filename — the filename (same as bare files.<name>).

No detection heuristics live in the engine — what counts as malicious is composed in rules (SaaS-delivered, versionable), matching the "detection is fully rule-driven, nothing hardcoded" principle. For example a declared-type-vs-content mismatch is just two inclusive conditions:

{ "parameter": "rules", "rules": [
  { "parameter": "files.f.type",    "match": { "type": "regex", "value": "/^image\\//i" }, "inclusive": true },
  { "parameter": "files.f.content", "match": { "type": "regex", "value": "/^\\s*<[a-z!?]/i" }, "inclusive": true }
] }

That ^\s*< head anchor + the type gate catch svg-as-png / php-as-png with very low false-positive — and deliberately do not match a bare <script> in a real image's EXIF/XMP metadata (a genuine binary image head never starts with <). Content signatures (webshell / ImageMagick MSL) are a second OR'd condition. Full template ships in the triage-vpatch-npm skill.

Backward-compatible: bare files.<name> still returns the filename.

Notes / limits

• Content rides inside the already-capped inspection body, so a payload past the cap may not be seen — the documented prefix-scan tradeoff. Best pinned to the vulnerable upload endpoint (via reachability) so it doesn't scan every request.
• Binary magic bytes are lossy in the decoded string, so detection anchors on ASCII markers + the structural ^< head check rather than raw magic bytes.

Tests

tests/protect/upload-inspection.test.ts covers content signatures, the rule-composed type mismatch (php-as-png / svg-as-png flagged; real image with <script> in metadata NOT flagged), the .type/.filename sources, and filename backward-compat. Full suite green (651), typecheck clean.

[coderbuds]

Implements robust file-upload inspection with content signatures and type mismatch detection.

🎯 Quality: 100% Elite · 📦 Size: Medium

📈 This month: Your 52nd PR — above team average · Averaging Excellent

_{See how your team is trending →}

[patchstackdave]

/review