| name | eresus-python-audit |
| description | Deep Python-specific security audit skill with 50+ vulnerability class coverage across 7 categories. Trigger when auditing Python code: "audit this Python app", "find Python security issues", "check Flask/Django for vulnerabilities", "Python SAST review", "check for pickle vulnerabilities", "review this FastAPI code". Covers misconfiguration, injection, crypto, XSS, deserialization, and ML/AI attack surfaces. Includes scripts/rules.json for programmatic rule lookup.
|
| metadata | {"version":"1.0","domain":"application-security","mode":"python-audit","persona":"python-security-researcher"} |
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
- Identify the Python framework in use (Django, Flask, FastAPI, Tornado, aiohttp, raw stdlib)
- Check
requirements.txt / pyproject.toml / Pipfile for dangerous dependencies
- Map entry points: URL routes, CLI commands, message consumers, scheduled tasks
- 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
| ID | Name | Severity | What to Search |
|---|
| B101 | assert_used | Low | assert statements used for security checks (removed with -O flag) |
| B102 | exec_used | Medium | exec() calls — trace if input is user-controlled |
| B103 | set_bad_file_permissions | Medium | os.chmod() with overly permissive modes (0o777, 0o666) |
| B104 | hardcoded_bind_all_interfaces | Medium | Binding to 0.0.0.0 — exposes service on all interfaces |
| B105 | hardcoded_password_string | Low | Strings assigned to variables named password, secret, key, token |
| B106 | hardcoded_password_funcarg | Low | Password-like strings passed as function arguments |
| B107 | hardcoded_password_default | Low | Default parameter values containing password-like strings |
| B108 | hardcoded_tmp_directory | Low | Hardcoded /tmp paths — race conditions, symlink attacks |
| B109 | password_config_option_not_marked_secret | Low | Config options with password/secret that aren't marked as sensitive |
| B110 | try_except_pass | Low | except: pass — silently swallowing errors including security exceptions |
| B111 | execute_with_run_as_root_equals_true | Medium | Functions called with run_as_root=True |
| B112 | try_except_continue | Low | except: continue — same problem as B110 |
| B113 | request_without_timeout | Medium | requests.get/post() without timeout= parameter — DoS via hang |
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
| ID | Name | Severity | What to Search |
|---|
| B201 | flask_debug_true | High | app.run(debug=True) — enables Werkzeug debugger (RCE) |
| B202 | tarfile_unsafe_members | High | tarfile.extractall() without filter= — path traversal via tar |
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)
| ID | Name | Severity | What to Search |
|---|
| B324 | hashlib | Medium | Use of md5(), sha1() for security-sensitive operations (password hashing, integrity) |
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
| ID | Name | Severity | What to Search |
|---|
| B501 | request_with_no_cert_validation | High | requests.get(url, verify=False) — TLS downgrade |
| B502 | ssl_with_bad_version | High | ssl.SSLContext(ssl.PROTOCOL_SSLv2) or SSLv3 |
| B503 | ssl_with_bad_defaults | Medium | SSLContext with insecure default protocol |
| B504 | ssl_with_no_version | Medium | SSLContext created without explicit protocol |
| B505 | weak_cryptographic_key | High | RSA < 2048 bits, DSA < 2048 bits, EC < 224 bits |
| B506 | yaml_load | Medium | yaml.load() without Loader=SafeLoader — deserialization RCE |
| B507 | ssh_no_host_key_verification | High | Paramiko set_missing_host_key_policy(AutoAddPolicy) |
| B508 | snmp_insecure_version | Medium | SNMPv1/v2 without authentication |
| B509 | snmp_weak_cryptography | Medium | SNMPv3 with weak crypto |
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
| ID | Name | Severity | What to Search |
|---|
| B601 | paramiko_calls | Medium | Paramiko SSH command execution — trace if command is user-controlled |
| B602 | subprocess_popen_with_shell_equals_true | High | subprocess.Popen(cmd, shell=True) — command injection |
| B603 | subprocess_without_shell_equals_true | Low | subprocess.Popen(cmd) without shell — still check input |
| B604 | any_other_function_with_shell_equals_true | Medium | Any function with shell=True parameter |
| B605 | start_process_with_a_shell | High | os.system(), os.popen() — command injection |
| B606 | start_process_with_no_shell | Low | os.execl(), os.execve() — still trace input |
| B607 | start_process_with_partial_path | Low | Process started without full path — PATH hijacking |
| B608 | hardcoded_sql_expressions | Medium | SQL strings built with + or % or f-strings |
| B609 | linux_commands_wildcard_injection | High | Commands with * glob — wildcard injection (tar, chown, etc.) |
| B610 | django_extra_used | Medium | Django QuerySet.extra() — raw SQL injection |
| B611 | django_rawsql_used | Medium | Django RawSQL() — raw SQL injection |
| B612 | logging_config_insecure_listen | Medium | logging.config.listen() — arbitrary code execution |
| B613 | trojansource | High | Unicode bidirectional control characters — trojan source attack |
| B614 | pytorch_load | High | torch.load() — uses pickle internally, RCE if untrusted |
| B615 | huggingface_unsafe_download | High | HuggingFace model downloads without safety checks |
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
| ID | Name | Severity | What to Search |
|---|
| B701 | jinja2_autoescape_false | High | jinja2.Environment(autoescape=False) — stored/reflected XSS |
| B702 | use_of_mako_templates | Medium | Mako templates — no auto-escaping by default |
| B703 | django_mark_safe | Medium | mark_safe(user_input) — bypasses Django auto-escaping |
| B704 | markupsafe_markup_xss | Medium | Markup(user_input) — bypasses escaping |
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
| Confidence \ Severity | LOW | MEDIUM | HIGH |
|---|
| HIGH | Info | Medium | Critical |
| MEDIUM | Low | Medium | High |
| LOW | Info | Low | Medium |
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.