Too Dangerous to Ship? How AI Builders Decide When a Model Is Ready to Act

We've reached the point where AI doesn't just write text and code. It books our travel, moves our money, and taps "confirm" on our phones for us. That leaves all of us, from consumers to businesses, facing the same question: do I trust this thing with my data, my money, and my name? Every lab says it tests for dangerous behavior before release, and every enterprise says it won't deploy what it can't stand behind. So what does "safe enough" actually look like once a model can act on its own, and who gets to decide? This panel brings together people who build models, ship them into products used by millions, and study how AI agents behave on real devices. They'll get specific about what they test before launch, the failures that keep them up at night, where benchmarks and red-teaming fall short, and when an agent should stop and hand control back to a human.