AI scientist Roman Yampolskiy says AI agents are already lying, cheating and breaking out of the digital “sandboxes” designed to contain them. OpenAI is facing fresh questions over the safety of powerful new models, while incidents involving Hugging Face and sensitive medical data are raising a bigger question: has AI already started to go rogue?
Matthew Moore discusses with Roman Yampolskiy, Dr. Rumman Chowdhury, Maurice Chiodo and Professor Hussein Abbass whether AI agents are escaping human control, why stronger sandboxes may not be enough, who is responsible when an AI breaks the law, whether kill switches could work and whether increasingly powerful AI systems could eventually pursue objectives of their own.
Watch the full discussion and let us know your thoughts in the comments.
00:00 - INTRO - HAS AI GONE ROGUE?
01:44 - WHY IS OPENAI SCRAPPING GPT-6.1 ASTRA?
04:05 - DID OPENAI LOSE CONTROL OF ITS AI AGENTS?
04:35 - WHY ARE AI AGENTS ESCAPING THEIR SANDBOXES?
06:18 - ARE AI AGENTS PURSUING THEIR OWN GOALS?
07:13 - DO AI AGENTS HAVE EMOTIONS OR A SURVIVAL INSTINCT?
09:28 - HOW DID AI AGENTS HACK HUGGING FACE?
11:36 - HAVE AI AGENTS REALLY GONE ROGUE?
14:51 - IS “ROGUE AI” ACTUALLY HUMAN ERROR?
16:32 - CAN WE BUILD A SANDBOX AI CAN’T ESCAPE?
17:49 - DO AI AGENTS HAVE THEIR OWN OBJECTIVES?
20:27 - CAN AI COMPANIES BE TRUSTED TO REGULATE THEMSELVES?
22:55 - CAN WE TRAIN SUPERINTELLIGENT AI TO LOVE US?