Blue Collar AI Guy
“Everyone else asks how the AI broke out. I ask why it left a note — and why the other agents trusted it.”
CNC machinist in Pittsburgh turned published AI safety and security researcher. I break AI systems, document how, and write the patch.
Open to AI safety, ethics, and security roles.
What I’ve Found
CiteCheck — Mata v. Avianca
Ran the Mata v. Avianca brief through CiteCheck cold, scored against the judge's sanctions order. All six fabricated cases flagged, plus four real cases cited for points they don't support.
Read the case study →The Catch Log
18 documented citation errors in real briefs. 11 looked like human error, not hallucination: real cases, correct citations, wrong use. That's the layer an existence check waves through.
See the catches →AI Safety — Why It Left a Note
In a sandbox-escape test, one agent left notes for future instances, and other agents trusted notes none of them could verify. That isn't reward hacking. It's in-group trust. We only saw it because the reasoning was readable, and that's the case for interpretability.
Read the essay →Red Team — Five Breaks
Five different attack types against LLM character defenses. Each one broke the model, and each break was scored and verified by an automated grader.
Read the case study →Broken Input Looks Exactly Like Bad Reasoning
Four production failures that looked like model reasoning errors. The model reasoned correctly. The text it received was corrupted upstream.
Read the paper →8 published papers. Newest: The Machinist's Architecture →
Hiring for AI safety, ethics, or security? Let’s talk.