tanviet12/vbsec / skills/antigravity/vbs-scan-security
vbs-scan-security
FlaggedUse when scanning code for security vulnerabilities. Use when user says "scan security", "kiểm tra bảo mật", "security audit", "review security", or invokes `/vbs-scan-security`. For large scans (>20 main-language files OR >30 total OR >14 days) processes chunks sequentially. Outputs bilingual reports (vi/en). Optional `--auto-fix` (agentic patch + verify loop) and `--sca` (live CVE lookup via OSV.dev).
rules/languages/python/08-insecure-deserialization.md
View on GitHub---
id: INSECURE-DESERIALIZATION
severity_max: CRITICAL
applies_to: python
---
# Insecure Deserialization — Python Specialization
> Override cho rule chung `rules/generic/08-insecure-deserialization.md`. Áp dụng khi `primary_language: python`.
## Intent (Python-specific)
Python có **lớp lỗ hổng đặc biệt nguy hiểm hơn các ngôn ngữ khác**: `pickle.loads()` cho phép arbitrary code execution **by design** — chính Python docs có cảnh báo "WARNING: Never unpickle data received from an untrusted source". Cùng họ: `marshal`, `shelve`, `dill`, `jsonpickle`. `yaml.load()` không kèm `SafeLoader` cũng cho phép RCE qua `!!python/object/apply:os.system`. Vibe code rất hay dùng pickle để cache, session, queue (Celery), ML model — và load lại từ Redis/file/network mà không verify nguồn.
## Khi nào CRITICAL
- `pickle.loads(data)` / `pickle.load(file)` với data từ:
- HTTP request (L1)
- Redis/Memcached/queue không kiểm soát producer
- File user upload
- Cookie session (Django pre-default-JSON, Flask `pickle` session)
- `yaml.load(data)` KHÔNG có `Loader=SafeLoader` hoặc `yaml.safe_load`
- `marshal.loads`, `marshal.load` với input untrusted
- `dill.loads` / `dill.load` (siêu tập của pickle, còn permissive hơn)
- `jsonpickle.decode(data)` — mặc định cho phép `py/object` reference class
- `shelve.open(user_controlled_path)` — backing store là pickle
- `eval(user_input)` / `exec(user_input)` — RCE trực tiếp (cũng cross-ref `21-command-injection`)
- `compile(user_input, ...) + exec` — biến tướng eval
- Celery `task_serializer = 'pickle'` (default đã JSON từ 4.0+ nhưng vibe code có thể bật pickle)
## Khi nào HIGH (giảm cấp)
- Input từ trusted internal source có HMAC/signature verify trước khi unpickle
- `pickle.loads` trong test fixture rõ ràng
- `yaml.load` với `Loader=SafeLoader` (đây thực ra là SAFE — nếu code đã đúng thì không flag)
## Cách reasoning
1. **Grep** sink: `pickle\.loads?`, `marshal\.loads?`, `yaml\.load\(`, `dill\.loads?`, `jsonpickle\.decode`, `shelve\.open`, `\beval\(`, `\bexec\(`
2. **Read** function chứa call. Source của bytes/string là gì?
- L1: `request.body`, `request.files`, `request.cookies`, request param
- L2: Redis key, file path từ DB
- L3: file config local
- L4: hardcoded bytes
3. **Verify**:
- `yaml.load` có `Loader=SafeLoader` hoặc `yaml.SafeLoader`?
- pickle có HMAC sign verify trước không (`hmac.compare_digest`)?
- Celery config có `accept_content = ['json']` không?
- Session backend là gì? (`SESSION_ENGINE` Django, `SESSION_TYPE` Flask)
## Search patterns
```
# Pickle - CRITICAL by default
pickle\.loads?\s*\(
cPickle\.loads?\s*\(
_pickle\.loads?\s*\(
# Marshal - CRITICAL
marshal\.loads?\s*\(
# YAML không safe
yaml\.load\s*\((?![^)]*Loader\s*=\s*(yaml\.)?SafeLoader)
# (tìm yaml.load mà KHÔNG kèm SafeLoader)
# dill/jsonpickle
dill\.loads?\s*\(
jsonpickle\.decode\s*\(
# shelve
shelve\.open\s*\(
# eval/exec với khả năng L1
\beval\s*\(\s*(request\.|input\(|sys\.argv)
\bexec\s*\(\s*(request\.|input\(|sys\.argv)
compile\s*\([^)]*\)\s*\)\s*\)?[^;]*\bexec
__import__\s*\(\s*request\.
# Celery cấu hình pickle
task_serializer\s*=\s*["']pickle["']
accept_content\s*=\s*\[[^\]]*["']pickle["']
result_serializer\s*=\s*["']pickle["']
# Django session với pickle
SESSION_SERIALIZER\s*=\s*["'][^"']*PickleSerializer["']
# Flask session pickle hoặc itsdangerous unsafe
session_interface\s*=\s*[^J][^S][^O][^N]
```
## Examples
### CRITICAL — flag
```python
# Flask + pickle từ request body — RCE
@app.post('/import')
def import_data():
data = pickle.loads(request.data) # L1 → RCE
return jsonify(ok=True)
# Exploit payload tạo bằng pickle.dumps(obj với __reduce__ chạy os.system)
```
```python
# Django + yaml.load không safe
def upload_config(request):
config = yaml.load(request.FILES['config'].read()) # CRITICAL
# Attacker upload YAML: !!python/object/apply:os.system ["rm -rf /"]
return JsonResponse(config)
```
```python
# Cache pickle từ Redis không HMAC
def get_user_cache(user_id):
raw = redis.get(f"user:{user_id}") # L2 — kẻ tấn công ghi vào Redis
if raw:
return pickle.loads(raw) # RCE nếu Redis bị compromise
```
```python
# FastAPI eval từ request
@app.post('/calc')
async def calc(expr: str): # L1
return {"result": eval(expr)}
# Exploit: expr = "__import__('os').system('id')"
```
```python
# jsonpickle decode mặc định cho phép class
@app.post('/restore')
def restore():
obj = jsonpickle.decode(request.data.decode()) # py/object → RCE
return jsonify(state=str(obj))
```
```python
# Celery với pickle serializer
# celery_config.py
task_serializer = 'pickle' # CRITICAL nếu broker chia sẻ
accept_content = ['pickle', 'json']
# Bất kỳ ai publish vào queue đều có thể RCE worker
```
### NOT critical — không flag
```python
# yaml.safe_load — SAFE
data = yaml.safe_load(request.files['config'].read())
# Hoặc explicit SafeLoader
data = yaml.load(stream, Loader=yaml.SafeLoader)
```
```python
# ast.literal_eval thay cho eval — SAFE cho literal
import ast
result = ast.literal_eval(request.json['expr'])
# Chỉ parse được str/int/float/list/dict/tuple/None/bool
```
```python
# JSON serializer — SAFE
data = json.loads(request.data)
```
```python
# Pickle với HMAC verify — chấp nhận nếu key bí mật
data = request.data
sig = request.headers['X-Sig']
expected = hmac.new(SECRET, data, 'sha256').hexdigest()
if hmac.compare_digest(sig, expected):
obj = pickle.loads(data) # OK vì verify integrity
```
## Fix recommendation
1. **Thay pickle bằng JSON** cho mọi data đi qua network/storage không tin cậy:
```python
# BAD
pickle.dumps(obj) → pickle.loads(bytes)
# GOOD
json.dumps(obj.__dict__) → json.loads(s)
# Hoặc msgpack, protobuf, flatbuffers cho hiệu năng
```
2. **YAML** — luôn dùng `safe_load`:
```python
# BAD
yaml.load(stream)
# GOOD
yaml.safe_load(stream)
```
3. **eval/exec** — không bao giờ với input user. Thay bằng:
```python
import ast
ast.literal_eval(s) # cho literal
# Cho expression toán: dùng asteval, simpleeval (sandboxed)
```
4. **Celery** — bắt buộc JSON:
```python
task_serializer = 'json'
accept_content = ['json']
result_serializer = 'json'
```
5. **Django session** — KHÔNG dùng pickle serializer (default đã là JSON từ Django 1.6+):
```python
SESSION_SERIALIZER = 'django.contrib.sessions.serializers.JSONSerializer'
```
6. **Nếu BẮT BUỘC pickle** (vd: ML model giữa trusted services) — kèm HMAC:
```python
def safe_pickle_load(data: bytes, sig: bytes, key: bytes):
expected = hmac.new(key, data, 'sha256').digest()
if not hmac.compare_digest(sig, expected):
raise ValueError("Signature mismatch")
return pickle.loads(data)
```
7. **ML model loading**: nếu load model từ HuggingFace/internet → ưu tiên format `safetensors` thay pickle/torch `.pt`.
## Cross-references
- Python `21-command-injection`: `eval`/`exec`/`__import__` là biến thể RCE qua deserialization
- Generic `01-hardcoded-secret`: HMAC key cho pickle verify không được hardcode
- Generic `20-outdated-dependency`: PyYAML <5.1 có CVE-2017-18342, dill cũ có CVE
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 470k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 470k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 470k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |