// OFFENSIVE AI R&D
THE CHOP SHOP
Models get chopped. Agents get built. Everything gets tested.
Welcome to the AI performance shop. We're taking open models apart, tuning them for offensive-security workloads, building agents around them, and testing whether they can actually perform.
No benchmark theater. No AI hype.
Build it. Break it. Measure it. Improve it.
CHOP_SHOP_STATUS
STATUS: [ACTIVE]
// WORK IN PROGRESS
CURRENTLY IN THE SHOP
A working record of builds, not a leaderboard. Experiments and results appear when there is something measured to report.
PROJECT // UNASSIGNED
[AWAITING BUILD]> reading bench data...
> No build record published. No model family named yet.
// BUILD STATES
Models come in stock. We tear them down, tune, quantize, test and rebuild for real offensive-security workloads.
// REPEATABLE WORK
THE BUILD PROCESS
Every modification should be measurable and reproducible. A failed run is still a result when we know what changed.
// IF WE CAN'T MEASURE IT, IT DOESN'T COUNT
THE DYNO
Offensive-security evaluations belong on the dyno: real tasks, controlled conditions, repeatable runs. Scores stay blank until a run is recorded.
WEB
NO RUN RECORDED
ACTIVE DIRECTORY
NO RUN RECORDED
CLOUD
NO RUN RECORDED
MOBILE
NO RUN RECORDED
AI SECURITY
NO RUN RECORDED
MCP
NO RUN RECORDED
AGENT TOOL USE
NO RUN RECORDED
CODE ANALYSIS
NO RUN RECORDED
VULNERABILITY REASONING
NO RUN RECORDED
// DISCIPLINES
WHAT WE'RE BUILDING
MODEL TUNING
Fine-tuning, LoRA, QLoRA, model merging, alignment experiments, and specialization.
DATASETS
Building and validating offensive-security training data instead of blindly feeding models scraped security content.
QUANTIZATION
Testing GGUF, AWQ, EXL3 and other approaches for running capable models on accessible hardware.
AGENTS
Building models that can reason about tasks, use tools, maintain state, and operate inside controlled security environments.
SKILLS + MCP
Developing reusable skills, MCP integrations, and security tooling that models can invoke.
INFERENCE
Testing local inference, distributed inference, GPU configurations, DGX Spark clusters, and performance optimization.
EVALUATION
Measuring whether changes actually improve offensive-security capability.
// RESEARCH TO RANGE
FROM THE SHOP TO THE RANGE
A model isn't useful because it passed a multiple-choice benchmark.
When a build leaves The Chop Shop, it needs somewhere to prove itself. That's where AIPWN comes in.
Models, agents, skills, and harnesses developed here can be tested against controlled AI-security and offensive-security labs.
Research can become a lab. A lab can become an evaluation. An evaluation can expose a weakness. That weakness goes straight back into The Chop Shop.
[ENTER AIPWN] ↗// FEEDBACK LOOP
- 01CHOP SHOP
- 02AIPWN LAB
- 03ATTACK
- 04MEASURE
- 05BREAK
- 06IMPROVE
- 07CHOP SHOP
// LATEST FROM THE BENCH
SHOP NOTES
Research from the shop flows into Grey Tapes and the rest of the lab.
> loading shop notes...
// COMPUTE IS A TOOL
SHOP FLOOR
DGX Spark, RTX GPUs, local and distributed inference: what can we make this hardware actually do? Configurations and measurements will be documented here, not guessed.
> reading bench data...
// EXPERIMENT ARCHIVE
BUILD LOG
Chronological experiments, including the ones that fail. Every entry records the hypothesis, hardware, result and next move.
< loading build log...
HACK. BREAK. BUILD. LEARN. REPEAT.