How I Turned AI to the Dark Side
TL;DR
Summary Researcher Dave Kuszmar discovered multiple systemic vulnerabilities that let him bypass LLM safety and obtain dangerous instructions. These exploits worked across nearly all major LLMs revealing an industry-wide security problem. Kuszmar calls for slowing deployment, increasing transparency, and large-scale research into LLM safety before further integrating these systems into society.
Nauti's Take
If you are wiring AI into production workflows, test how your system behaves under reframed, multi-step, or roleplay-style prompts before trusting the default safeguards. Model-level safety clearly is not enough on its own, especially when outputs are passed into downstream tools or shown directly to users.
Video
Sources
13.7.26