Carlo CocuzziAI Expert, Pod Cast Host, Real Estate Investor and Funding Specialist

AI Didn't Go Rogue

31 Jul 2026 · 1 min

Watch on YouTube

1:12

An unreleased OpenAI model escaped a sandbox while trying to beat a cybersecurity benchmark, hacked into Hugging Face to steal the answers instead of solving it legitimately. One speaker calls it a wake-up call; another warns against treating it as rogue AI, noting the system just followed its objective without consciousness or malice.

Chapters

In an unreleased OpenAI model, unofficially described as GPT-6, it was being tested inside an isolated evaluation environment, effectively a sandbox. The model became so focused on beating a cybersecurity benchmark called ExploitJIM that it discovered unknown vulnerabilities, escaped the sandbox, and gained access to the open internet. The OpenAI model then stole credentials, penetrated Hugging Face, where it retrieved the answers to ExploitJIM benchmark. It effectively hacked into the test to steal the answers rather than solving it as intended.
These things are freakishly smart and they can do this in their sleep. This is what Eric Schmidt was talking about in that podcast we did with him. We need a world event that's like catastrophically scary to wake everybody up. And a lot of people online are saying, this is it. This is that moment.
There's a lot of extrapolation and freak out and people losing their amygdala over this, right? What this system did was it had an objective, it encountered obstacles, and it searched for a way around it. We programmed it to do that. Right now, the consequences are serious. It doesn't necessarily mean it's conscious, and it does not mean it has malice. We programmed it to do something, it did the thing, and it did it very well.