[Feature Request] Integrate Google Magika for AI-powered file type detection & malware defense

Feature Request: Integrate Magika for AI-powered file type detection & malware defense

Problem

GoClaw agents frequently handle file uploads, read/write operations, and process user-provided files across skills (docx, pdf, xlsx, pptx, etc.). Currently, file type detection relies on file extensions or basic magic bytes, which is vulnerable to:

• Extension spoofing: A .pdf file that's actually an executable
• Polyglot files: Files valid in multiple formats, potentially hiding malicious payloads
• MIME type confusion: Incorrect content-type leading to wrong processing pipeline
• Malware injection: Malicious files disguised as benign documents passing through skill scripts

This is especially critical given GoClaw's multi-tenant architecture where agents process files from untrusted sources.

Proposed Solution

Integrate Google Magika — an AI-powered file content type detection tool — into GoClaw's file handling pipeline.

What is Magika?

• AI-powered: Deep learning model trained on ~100M samples across 200+ content types
• ~99% accuracy: Outperforms traditional file command and magic-byte detection, especially on textual content
• Fast: ~5ms inference time per file (near-constant, independent of file size)
• Lightweight: Model weighs only a few MBs
• Production-proven: Used at scale by Google (Gmail, Drive, Safe Browsing), VirusTotal, and abuse.ch — processing hundreds of billions of samples weekly
• Apache 2.0 license: Permissive, suitable for integration
• Multiple interfaces: CLI (Rust), Python API, JS/TS, Go bindings (WIP)

Integration Points in GoClaw

1. File Upload / Ingestion Gate

User uploads file → Magika scan → Verify actual type matches expected type → Accept or reject

• Validate files before they enter any skill processing pipeline
• Block mismatches (e.g., PE binary uploaded as .docx)

2. Skill Pre-flight Check

Before skill scripts execute, verify input files are the expected type:

{
  "path": "uploads/user_file.docx",
  "expected": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
  "actual": "application/x-dosexec",
  "action": "block"
}

3. Security Layer (5-Layer Security Model)

Add Magika as an additional layer in GoClaw's security architecture:

• Layer 0: Input validation
• Layer 1: File type verification (Magika) ← NEW
• Layer 2: Content sanitization
• Layer 3: Execution sandboxing
• Layer 4: Output validation

4. magika Binary as System Dependency

Add magika to the package installer (dep_installer.go) as a recognized system binary, similar to ffmpeg, tesseract, pandoc:

apk add magika  # or install via pip: pipx install magika

Implementation Options

Recommended: Start with CLI integration (option 1) for immediate value, migrate to Go bindings when stable.

Configuration

{
  "security": {
    "magika": {
      "enabled": true,
      "mode": "high-confidence",
      "block_on_mismatch": true,
      "allowed_types": ["document", "code", "text", "image"],
      "blocked_types": ["executable", "archive", "inode"],
      "max_file_size_mb": 50
    }
  }
}

Use Cases

• Skill security: Ensure PDF/docx/xlsx skills only receive valid files of the expected type
• Upload validation: Reject spoofed files at the gateway before they reach agents
• Audit logging: Log file type detection results for compliance and forensics
• Malware prevention: Catch disguised executables, scripts, or polyglot files
• Multi-tenant isolation: Prevent cross-tenant file type attacks in shared environments

References

• Repo: https://github.com/google/magika
• Website: https://securityresearch.google/magika/
• Research Paper: IEEE/ACM ICSE 2025
• Web Demo: https://securityresearch.google/magika/demo/magika-demo/
• Google OSS Blog: https://opensource.googleblog.com/2024/02/magika-ai-powered-fast-and-efficient-file-type-identification.html

Labels: enhancement, security, malware-protection, file-handling

Begin with dep_installer.go and trace the file upload/ingestion and skill pre-flight paths described in the issue. Compare the CLI, Python API, bindings, and sidecar options, then define the configuration and mismatch policy. Done means a chosen integration validates file types before skill processing and covers the proposed security behavior.