taintlens-code-audit

taintlens-code-audit

A free, MIT-licensed security-audit skill for coding agents — Claude Code, DeepSeek Harness (dsh), Codex-style CLIs, Cursor, anything that reads the SKILL.md format.

It audits code the way an attacker reads it: map the entry points, follow attacker-controlled data, land on a sink — and only report a finding when the path is walkable, with the minimal request that proves it.

[SEV-High 8] SQLi — orders/api.py:57
Entry point : GET /orders/<id>
Taint path  : request.view_args["id"] -> f-string -> cursor.execute()
Sanitizer   : none (string interpolation, not parameter binding)
Authz       : session user checked, but row ownership not (IDOR also open)
PoC         : GET /orders/1%20OR%201=1 with session cookie of user B
Fix         : cursor.execute("... WHERE id=%s AND owner=%s", (id, session.user))

No walkable path means a Hardening appendix entry — never an inflated "vulnerability". That honesty rule is the whole point: an audit you cannot act on is noise.

What the skill contains

• A 7-pass method: entry points → sources → sinks → taint walk → per-object authz → secrets → dependencies
• A 0–10 severity rubric with explicit adjust rules, so two runs produce comparable bands
• Sink cheat-sheets for Python, JS/TS, Java, Go, PHP
• A strict output contract: file:line, taint path, sanitizer verdict, authz verdict, minimal PoC, smallest fix

Install

Copy this folder into whatever your agent treats as a skills directory:

• Claude Code: ~/.claude/skills/taintlens-code-audit/
• DeepSeek Harness: copy (or junction) into %USERPROFILE%\.dsh\skills\
• Anything else: point your agent at the folder and let it read SKILL.md

No dependencies, no scripts to run, nothing leaves your machine — it is a method written down, not a program. Python 3.14 wrote the example; the skill itself is just instructions.

Try it in two minutes

1. Install the skill (above).
2. Point your agent at examples/vulnerable/ and ask it to audit the app.
3. Compare its report with examples/expected-report.md.

Scan your skills too

Installing third-party skills is a supply-chain decision. Before you install this or any other skill, run it through a static gate: skill-firewall — python firewall.py scan ./taintlens-code-audit — zero dependencies, and this repo is configured to self-scan clean.

When a skill is not enough

An AI pass is triage, not an audit. If the repository faces real attackers — customer data, payments, production traffic — findings still need a human who has owned the blast radius. This skill grew out of that service:

TaintLens Studio — taint-based first pass (hours), human verification, every finding PoC-proven, for teams that need an audit they can act on.

Related

• dsh-plugin-hub — one control hub for plugins and skills in DeepSeek Harness
• skill-firewall — pre-install security gate for agent skills

MIT licensed. See LICENSE.