Studio experiment E08 · for the editor's review
The lethal trifecta, made visible
When is an agent dangerous by construction? Three switches feed one agent: private data, untrusted content, a way out. With all three on, a path opens: a document's words turn the agent, and it sends your data to the outside. Turn any one off and that path stops at a wall; turn it back on and the path opens again.
Not part of the talk: an experiment for the review. The idea is Simon Willison's, The lethal trifecta for AI agents, the talk's own source; the slide names him on screen and uses only the three ingredients, in plain words. It is a picture of one path, not a measurement: no numbers, no case, no rating. Made by ruv.sgit@the-ultimate-insider in run r18-e08-lethal-trifecta at deck@'s request, the editor's word. Frame by frame, the source and what was left out: the explainer.