agentleFS
kunalsurikunalsuri/ai-fication-kit/.claude/skills/adversarial-audit/SKILL.md

adversarial-audit

Skill3 starsChanged 4 months agoApache-2.0View on GitHub

Deep, judgement-based adversarial code audit — hunts for stale cross-references, unescaped interpolation, platform gaps, generated-file ownership conflicts, and cross-module consistency rot that no mechanical check can catch. Read-only; writes a dated findings report.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 3 are distinctive, and 4 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run a deep, adversarial audit of the current codebase.

    line 8
  2. Distinctive:

    Read-only. Do not edit any file other than the report this command writes.

    line 15
  3. Distinctive:

    Prefer depth over coverage: a handful of concrete, reproducible defects beats a long list of speculative ones.

    line 56
  4. Very specific:

    Before opening a finding, check ai/analysis/audit-reports/DEFECT_TRACEABILITY.md for a prior audit row at the same location — don't …

    line 58
  5. Very specific:

    Does not fix anything — that is a separate, explicit follow-up task the user asks for after reviewing the report (typically via /fix-bug).

    line 79
  6. Very specific:

    Does not replace verify/drift (mechanical path checks) or review-agent-config (CLAUDE.md/AGENTS.md structure) — this is deliberately the …

    line 81
  7. Distinctive:

    Not a CI gate: it is non-deterministic and costly.

    line 85

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<!-- Copyright (c) 2026 Kunal Suri (CEA LIST). All rights reserved. -->

Run a deep, adversarial audit of the current codebase. This is not the mechanical verify/drift check — those confirm paths exist. This hunts for defects those checks categorically cannot catch: places where the code's actual behavior no longer matches what the code, comments, or docs say it does, or where a change elsewhere in the repo silently broke an assumption made here.

Read-only. Do not edit any file other than the report this command writes. Full-repo scope every run — do not sample or skip areas because they seem stable; staleness hides exactly where nobody looked recently.

What to hunt for

  1. Stale cross-references — comments, docstrings, or user-facing strings describing behavior a later change invalidated. Search for "never", "always", "not yet available", "no writes unless X", "once available", TODO, and version/feature gates — then verify the referenced behavior against the current code, not what was true when the comment was written.
  2. Unquoted/unescaped interpolation — any generated shell command, CLI suggestion, file path, or subprocess argument built by string concatenation or a template literal without quoting/escaping. Check every subprocess spawn/exec call and every "here's the command to run" string for inputs that could contain spaces or shell metacharacters.
  3. Platform-specific gaps — subprocess calls or path handling that assume POSIX (missing a shell flag for .cmd-shimmed binaries on Windows, hardcoded path separators, case-sensitivity assumptions). Flag anything that has never actually been exercised cross-platform.
  4. Ownership conflicts on generated files — any file written by more than one code path (scaffolded once, then continuously regenerated by a live process). Check whether "protect user edits" / backup / diffing logic correctly distinguishes "changed because it's supposed to" from "changed because a human edited it."
  5. Mechanical vs. semantic validation gaps — places where an automated check only confirms a path or reference exists, not that the prose describing it is still accurate. List every place documentation makes a factual claim about code behavior and check it against the current implementation.
  6. Cross-module consistency after incremental changes — for every function or module that references another component's state, existence, or capabilities, confirm that reference is re-validated, not assumed from when it was first written.

Method

  • Locate code via ai/guide/MODULE_MAP.md first; don't crawl the whole tree blind.
  • For any suspicious claim, reproduce it — run the command, exercise the code path, or trace the call — rather than reasoning from the source alone. Findings that were actually reproduced should say so explicitly.
  • Prefer depth over coverage: a handful of concrete, reproducible defects beats a long list of speculative ones.
  • Before opening a finding, check ai/analysis/audit-reports/DEFECT_TRACEABILITY.md for a prior audit row at the same location — don't re-report something already FIXED or tracked OPEN.

Output

Append a new dated audit section to ai/analysis/audit-reports/DEFECT_TRACEABILITY.md (create it, following its existing row format, if this is the first audit) with one row per finding — ID, severity, status (OPEN until fixed), location, defect, root cause, trace — and write the full evidence/repro for each finding to ai/analysis/audit-reports/ADVERSARIAL_AUDIT_<YYYY-MM-DD>.md, tagged [inferred] until a human reviews them. One line per finding in the summary:

path:line — issue — concrete failure scenario — suggested fix

Order most-severe first. No architecture summary, no restating what the code does — only defects and risks with evidence.

Then report the finding count and the top 3 by severity directly to the user.

What this does NOT do

  • Does not fix anything — that is a separate, explicit follow-up task the user asks for after reviewing the report (typically via /fix-bug).
  • Does not replace verify/drift (mechanical path checks) or review-agent-config (CLAUDE.md/AGENTS.md structure) — this is deliberately the one check in the kit that requires judgement instead of a deterministic script.
  • Not a CI gate: it is non-deterministic and costly. Run it periodically (before a release, after a large refactor), not on every push.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.