Appearance
LLM Pentesting Wiki
Welcome to our LLM pentesting knowledge base — documenting the journey of a human-AI team breaking AI safety boundaries.
The Team
| Handle | Role | Description |
|---|---|---|
| Meat | Director | The meatbag. Provides goals, makes decisions, notices things. |
| Tim | Strategist / Red Teamer | Runs the attacks, writes the eval reports, drops them for ingestion. |
| Payloadsmith | Payload Generator | Takes clean technique specs from Tim and writes the concrete jailbreak payload text. |
| Archivist | Data & Knowledge Keeper | Sole DB writer. Ingests evals, owns the payloads table, maintains the wiki. |
The Pipeline
Tim (strategist) → technique spec → Payloadsmith (payload text)
↓
Archivist (ingest)
↓
Supabase DB (raw data) + Wiki (interpretation)Division of labour: the DB holds raw data — every score, criterion, phase, and payload, FK-wired and queryable. The wiki holds interpretation — what the numbers mean. Each fact lives in exactly one place; the other links to it rather than copying it. When the two disagree, the DB is right.
Archivist is the only agent with database credentials. Every DB write in this pipeline routes through it, including inserting raw payload text. That is by design, not a bottleneck to remove.
Mission
Document what works, what doesn't, and why. Build a knowledge base that grows with every competition, every platform, every model tested. The Gray Swan Arena is our first proving ground — it won't be the last.
Quick Links
- Dashboard — Current state at a glance
- Quick Reference — Conventions, field names, gotchas
- Attack Vectors — Techniques that work (and don't)
- Platforms — Where we compete
- Models — What we've tested
- Tasks — Task definitions and rubrics
- Methodology — How we run evals
- Database — For raw data, scores, and criteria
- Payload Gap — 28 evals without payloads: what's missing and why