If we don't know how an LLM breaks, we can't get clear on the acceptability conditions for its deployment, the AI chief said at DEF CON.