Python Security Audit

Purpose

Perform a comprehensive, depth-first security audit of Python codebases. This skill provides the complete knowledge of Bandit's 50+ security checks, organized by category and severity, plus framework-specific patterns for Django, Flask, FastAPI, and emerging ML/AI attack surfaces.

Use view_file and grep_search exclusively. No terminal commands.

Audit Workflow

Phase 1: Reconnaissance

1. Identify the Python framework in use (Django, Flask, FastAPI, Tornado, aiohttp, raw stdlib)
2. Check requirements.txt / pyproject.toml / Pipfile for dangerous dependencies
3. Map entry points: URL routes, CLI commands, message consumers, scheduled tasks
4. Identify configuration files and secrets management approach

Phase 2: Systematic Check — By Category

Work through each category below. For each check, use grep_search to find all instances, then view_file to trace the data flow and confirm exploitability.

B1xx — Miscellaneous Checks

Audit Depth for B1xx

• B101: Check if assert guards authentication or authorization. If so, HIGH severity.
• B102: Trace exec() input — if user-controlled, escalate to CRITICAL (RCE).
• B105/106/107: Check if the hardcoded credentials are for production systems or test fixtures.
• B113: Check all HTTP client calls (requests, httpx, urllib3, aiohttp) for timeout.

B2xx — Application/Framework Misconfiguration

Audit Depth for B2xx

• B201: Check if debug=True is conditional on environment or always on. Check for WERKZEUG_DEBUG_PIN.
• B202: Any tarfile.open() + extractall() from user-uploaded files = CRITICAL path traversal.

B3xx — Dangerous Function Calls (Blacklists)

Extended B3xx Checks (Eresus additions)

• hashlib.md5() / hashlib.sha1() for password storage → escalate to HIGH
• Use of random module instead of secrets for security tokens → HIGH
• string.Template with user input → potential template injection

B5xx — Cryptography

Audit Depth for B5xx

• B501: Check if cert pinning is used. If verify=False is in production code → CRITICAL.
• B506: This is a deserialization sink. If input comes from HTTP/file upload → CRITICAL RCE.

B6xx — Injection

Audit Depth for B6xx

• B602/B605: Trace the command string backwards. If ANY part is user-controlled → CRITICAL RCE.
• B608: Check if the SQL is parameterized elsewhere. Raw f-string SQL = HIGH SQLi.
• B609: tar cf archive.tar * in /tmp with user-created files → argument injection.
• B614: torch.load() from user-uploaded model file → CRITICAL RCE via pickle.
• B615: ML model supply chain attack — check if trust_remote_code=True.

B7xx — XSS / Template Injection

Audit Depth for B7xx

• B701: If autoescape=False and template renders user input → CRITICAL XSS.
• B703: Trace what data is passed to mark_safe(). If user-controlled → HIGH XSS.

Framework-Specific Deep Checks

Django

• Check ALLOWED_HOSTS configuration (empty = open redirect)
• Check CSRF_COOKIE_HTTPONLY, SESSION_COOKIE_SECURE, SECURE_BROWSER_XSS_FILTER
• Check for @csrf_exempt decorators on sensitive views
• Check DEBUG = True in production settings
• Check SECRET_KEY hardcoded or in version control
• Check MIDDLEWARE ordering (SecurityMiddleware should be first)
• Check AUTH_PASSWORD_VALIDATORS configuration

Flask

• Check SECRET_KEY generation (must be cryptographically random)
• Check app.run(debug=True) in production
• Check for @app.before_request authentication enforcement
• Check session cookie configuration (SESSION_COOKIE_SECURE, SESSION_COOKIE_HTTPONLY)
• Check file upload handling (werkzeug.utils.secure_filename)

FastAPI

• Check for missing input validation (Body(), Query(), Path() without constraints)
• Check CORS configuration (allow_origins=["*"])
• Check authentication dependency injection (missing Depends())
• Check FileResponse / StreamingResponse path traversal
• Check OAuth2 implementation (token validation, scope enforcement)

ML/AI Attack Surface

• torch.load() — pickle-based, RCE from untrusted models
• transformers.pipeline(trust_remote_code=True) — arbitrary code execution
• pickle.loads() in model serialization pipelines
• Prompt injection in LLM-based applications
• Model poisoning via untrusted training data

Severity Matrix

Report Format

For each finding, report:

### [B-ID]: [Check Name]

**Severity**: [LOW/MEDIUM/HIGH/CRITICAL]
**Confidence**: [LOW/MEDIUM/HIGH]
**File**: [path]:[line]

**Vulnerable Code**:
[show the code]

**Data Flow**:
[source] → [intermediaries] → [sink]

**Impact**: [what an attacker achieves]
**Remediation**: [specific fix with code example]
**Bandit Reference**: B[xxx]

Tooling Constraints

Use ONLY:

• view_file — read source code
• grep_search — find patterns across the codebase

Do NOT use any terminal commands.