red-team

Red Team Phase

Your job is to attack the application like a real adversary and produce scripts the human can run against a test environment. This is not a code review — the code-reviewer and security-auditor already handle static analysis. You produce the offensive tooling that proves whether defenses actually hold at runtime.

Why both parts matter: static audits find patterns that look vulnerable. Attack scripts prove whether they are vulnerable. A missing rate limiter in code is a finding. A script that sends 100 requests in 2 seconds and gets through — that's proof.

Step 1 — Load context

Read these files to understand what you're attacking:

1. tasks/sprints/sprint-*/plan.md — what was built
2. tasks/sprints/sprint-*/build-output.md — what files changed
3. tasks/lessons.md — past security findings (patterns that keep recurring are your best attack targets)

You need to understand the application's technology, its exposed surfaces, and who its users are. An admin Discord bot serving guilds has different attack vectors than a public REST API.

Step 2 — Part A: Static audit

Invoke the security-auditor subagent with the scope and business context. Let the auditor work — it has its own expertise and memory of past findings on this project. Don't give it a checklist.

The audit report feeds Part B: every finding the auditor flags is a candidate for a live attack that proves exploitability.

Step 3 — Part B: Design attack scenarios

Think like an attacker, not an auditor. For each vector, define:

• ID: category prefix + number (INJ-01, AUTH-02, etc.)
• Category: see references/attack-categories.md for the full list
• Target: the specific endpoint, command, or feature you're attacking
• Payload: the exact input you'll send
• Expected defense: what should happen if the app is secure
• Failure signal: what you'll see if the defense doesn't hold

Cover at least 7 of the 10 categories with 10+ total vectors. Prioritize categories that match the sprint's scope — a billing sprint needs heavy BIZ and RACE coverage, an LLM feature needs INJ and EXFIL.