Research-First Discipline

Rule: NO payload leaves your hands without consulting the knowledge base.

Why This Exists

AI agents waste 40-60% of tool calls reinventing known techniques. The knowledge base (references/vuln-classes/ and references/chains/) contains field-tested payloads, bypass patterns, and chain recipes that cover the common cases. Skipping this step means slower engagements, weaker findings, and missed chains.

Protocol

Before Every Hunt

1. Identify the vuln class you're about to test (e.g., xss, sqli, ssrf).
2. Read the cheatsheet: references/vuln-classes/{class}.md
   • Detection payloads (what to try first)
   • Bypass patterns per WAF/framework
   • Validation criteria (when is it real)
   • Common mistakes for this class
3. Read chain patterns: references/chains/chain-patterns.md
   • What chains from this vuln class
   • What chains INTO this vuln class
   • Known escalation paths
4. Check AppProfile (engine/app_profile.py):
   • What tech stack? → Use framework-specific payloads
   • What WAF/CDN? → Use known bypass patterns
   • What auth mechanism? → Adapt to session handling
5. Check Brain (engine/brain.py):
   • What capabilities do we already have? → Skip redundant probes
   • What was already tried? → Don't repeat exhausted vectors
   • What vectors are active? → Coordinate, don't collide

Payload Selection Order

1. Known-good payload from cheatsheet for this tech stack
2. Variant of known-good adapted to observed WAF/encoding
3. Novel payload (only after 1 and 2 exhausted)

Never start with novel payloads. The cheatsheet payloads have been field- tested; novel payloads haven't.

After Every Finding

1. Record the working payload back to the Brain
2. Note the bypass if any WAF/filter was circumvented
3. Check chain anchors in the hunt skill's Chain Anchors section
4. Update capabilities in the Brain for chain engine

Validation Gate

Before any hunt agent fires its first payload:

□ vuln-class cheatsheet loaded? → if not, load it
□ AppProfile consulted? → if not, read it
□ Brain checked for prior work? → if not, check it
□ Payload from cheatsheet, not invented? → if novel, justify why

Failing this gate means the agent skipped research. The engagement controller SHOULD reject the action and force a research pass.

Gotchas

• Cheatsheets are starting points, not exhaustive. If the target uses an unusual framework, the agent MUST adapt — but start from the closest known pattern.
• Bypass patterns age fast. If a known bypass fails, mark it exhausted in the Brain and try the next one.
• Don't cargo-cult payloads. A cheatsheet payload for PHP won't work on Node.js SSTI. Match the payload to the confirmed tech stack.