agent-virtual-office: Skill for Codex

Red Team / Adversarial Testing

Overview

This skill applies adversarial thinking to code changes: instead of checking against a compliance list (that's what security_guardrails.md does), it actively asks "how would an attacker exploit this change?" and "what breaks under extreme conditions?"

It complements — never replaces — the existing OWASP security scan in /review.

Ironclad Rules

1. No bypass of governance: This skill executes within /review and /test phases only. It cannot override gates, skip phases, or alter classification.
2. Severity honesty: Only mark CRITICAL when there is a concrete, exploitable attack path with evidence (file:line). Speculative risks are HIGH at most.
3. Additive only: Red Team findings supplement existing security findings — never contradict or override them.

When to Use (Auto-Trigger Matrix)

AI MUST check the task classification from the Work Log and apply this matrix automatically:

Classification        │ /review          │ /test
──────────────────────┼──────────────────┼─────────────────
tiny-fix              │ —                │ —
quick-win             │ —                │ —
hotfix                │ Lite Red Team    │ Lite Adversarial (1-2 cases)
feature               │ Full Red Team    │ Adversarial Cases
architecture-change   │ Full Red Team    │ Adversarial Cases + Beast Mode

Auto-trigger logic: During /review or /test, read Classification: from the active Work Log. If classification is hotfix, feature, or architecture-change, execute the corresponding mode below. No user action required.

Modes

Lite Red Team (hotfix)

Minimal overhead (≤300 tokens output). Focus exclusively on the fix point:

1. Fix-Point Attack Vector: Does the fix itself introduce a new attack surface? (e.g., a validation fix that changes error behavior, a permission fix that alters fallback logic)
2. Regression Attack: Could an attacker exploit the old behavior to bypass the fix? (e.g., race condition in the fix window)