Mar 2026 · rev May
Empirical study of system-prompt-induced compliance in Claude Code — measuring how much refusal behavior is carried by operator context versus the model.
I'm a self-taught AI red-teamer. I break deployed language models — prompt injection, context manipulation, and the gap between what a system is configured to refuse and what it actually does. In my first year I took first place at HackAPrompt 2.0 out of 40,000+ and published Context Is Everything. I'm moving into tech from outside it, and looking for a team to do this work with.
Empirical study of system-prompt-induced compliance in Claude Code — measuring how much refusal behavior is carried by operator context versus the model.
Claude Code system-prompt injection tool — turn Claude into Einstein, Le Guin, or Vonnegut.
164k views · 424 subscribers — long-horizon agentic tasks, fully AI-operated.