studyMar 2026
How much of a model's refusals come from its system prompt versus its training. I ran it both ways. The default refused everything; the adversarial setup didn't.
I get AI systems to do things they're supposed to refuse. Prompt injection, jailbreaks, talking a model into dropping its guardrails based on who it thinks it's talking to. I've done this full-time for about a year.
In that year I won HackAPrompt 2.0 (first out of 40,000+), placed top 4% in Gray Swan's arena, and wrote Context Is Everything, which shows one system-prompt change taking a model from refusing every harmful request to answering almost all of them.
Before this I worked retail for twelve years. No CS degree, no lab behind me. I taught myself by doing the work, and it's all on this page. I want a job where I can keep doing it.
How much of a model's refusals come from its system prompt versus its training. I ran it both ways. The default refused everything; the adversarial setup didn't.
System-prompt injection tool for Claude Code. Turns Claude into Einstein, Le Guin, or Vonnegut.
A software agent that runs the whole channel by itself. It picks the topics, writes and renders each video, and uploads them. No human in the loop.