What "Leaked System Prompts" Actually Mean
Anthropic is the only major lab publishing its user-facing system prompts, yet the published versions leave out tool descriptions, accessory models, and orchestration. So when someone claims a prompt "leaked," are they showing you a real extraction or a model's confabulation? This episode traces the history of system prompt extraction — from Kevin Liu's one-sentence Bing attack in February 2023 to autonomous code agents recovering prompts from 41 commercial models — and looks at the defenses vendors built in response. The framing is defensive: understand the attack so you can defend against…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.