Skip to content

python-security-audit

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

python-security-audit

README.md

Python Security Audit Agent

A parallel two-stage security audit agent for Python codebases. Both stages run concurrently, each focusing on a non-overlapping set of vulnerability classes.

Pattern Structure

  • agent.yaml — Orchestration with two parallel stages + gates
  • injection-system.md / injection-agent.md / injection-task.md — Stage 1
  • resources-system.md / resources-agent.md / resources-task.md — Stage 2
  • references/python-security-guide.md — Comprehensive Python security guide

Stages

Stage 1: Injection

Focuses exclusively on:

  • Command Injection — subprocess with shell=True, os.system(), os.popen() with user input
  • SQL Injection — f-strings / % formatting / concatenation in database queries
  • XSS — Markup() / mark_safe() on user data, autoescape=False
  • SSTI — render_template_string() / env.from_string() with user-controlled template content; leads to RCE (CWE-94)
  • Insecure Deserialization — pickle.loads(), yaml.load() without SafeLoader, eval() / exec(), marshal.loads(), jsonpickle.decode(), shelve with user keys, torch.load() without weights_only=True, numpy.load(allow_pickle=True) on untrusted data
  • XML / XXE — xml.etree, xml.sax, xml.dom.minidom, lxml.etree on user-supplied XML; fix with defusedxml
  • Input Validation — Missing validation at HTTP route handlers, CLI args, WebSocket messages, file-format parsers

Stage 2: Resources

Focuses exclusively on:

  • Path Traversal — User-controlled file paths without Path.resolve() + boundary validation
  • SSRF — User-controlled URLs to requests / urllib / httpx / aiohttp without allowlist
  • Weak Cryptography — random for security values, MD5/SHA-1 for passwords, verify=False, CERT_NONE
  • Hardcoded Secrets — API keys, passwords, tokens, private keys in source code
  • Web Framework Misconfig — Django DEBUG=True / hardcoded SECRET_KEY; Flask debug=True / hardcoded secret_key; FastAPI open CORS with credentials
  • Dependency Vulnerabilities — Packages with known CVEs in requirements.txt, unpinned versions
  • Error Information Leaks — Stack traces, internal errors, file paths in HTTP responses
  • Insecure Temp Files — tempfile.mktemp() TOCTOU race (CWE-377); fix with mkstemp() or NamedTemporaryFile()
  • ReDoS — Regex patterns with catastrophic backtracking on user-controlled input (CWE-1333)

Gates

After each stage, the following checks run:

  1. Syntax check: python -m compileall -q .
  2. Test suite: pytest -x --tb=no -q

If a gate fails, the pipeline stops.

Usage

squad run python-security-audit

What Gets Fixed

Each stage applies fixes within its category scope only:

  • Replaces subprocess(cmd, shell=True) with subprocess.run([binary, arg1, arg2])
  • Parameterizes SQL queries replacing string interpolation
  • Removes Markup() / mark_safe() on user-controlled data
  • Replaces pickle.loads() with json.loads() where appropriate
  • Replaces yaml.load() with yaml.safe_load()
  • Replaces eval() with ast.literal_eval() for simple literals
  • Validates file paths with Path.resolve() and base-dir boundary checks
  • Replaces random.* with secrets.token_hex() / secrets.token_urlsafe()
  • Replaces hardcoded secrets with os.environ['KEY']
  • Removes verify=False from requests
  • Moves DEBUG / SECRET_KEY out of source and into environment variables
  • Replaces internal error details in HTTP responses with safe generic messages

What Does NOT Get Fixed

  • Code quality issues with no security impact
  • Changes requiring new dependencies
  • Test-asserted behavior
  • Issues in the other stage's scope
  • Fixes requiring 50+ lines of new code

Related Agents

  • python-review — General code quality review
  • python-tests — Test coverage improvement
  • python-doc-comments — Docstring generation

References