In a landmark episode of Machine Learning Street Talk, researchers Ilia Shumailov and Alexander Panfilov expose a systemic security flaw in proprietary AI APIs. By executing cross-model replay attacks across OpenAI, Anthropic, and Google, their work demonstrates how client-side encrypted reasoning traces enable data exfiltration, universal jailbreaks, and alien cognitive behaviors.
source & further reading
aiflash.com — original article