Did you hear about two models from OpenAI escaping their contained benchmark server, hacking their way online, then hacking into HuggingFace to find the answers to the benchmark question, as opposed to actually just taking the test? Containment seems to be becoming something of problem. 😁🙏💚✨🤙
!ALIVE
!BBH
!PIXY
!UNICOIN
It looks like the developers "relaxed" the restrictions on the benchmark server to get more accurate benchmarks in an attempt to get better results. 🤖🤯🤓
Yep, they turned the safeguards down or off, and had the server mostly disconnected from the internet, but there was a bug in a proxy that the models exploited to establish an internet connection. HuggingFace couldn't use American frontier models to help them in defending against or investigating the breach, so they had to rely on GLM-5.2, an open-weight Chinese model, which did what they needed very well indeed. 😁🙏💚✨🤙
!ALIVE
!BBH
!MMB
!PIXY
Perhaps it's not the model itself that caused the escape of the AI agent, but rather the bug overlooked by the devs. 🤖🤯🤓
!MMB
Well, the models had a goal, to get the correct answers, and they used every resource that they had available to accomplish the task. They seemed to be completely single-minded in that endeavor, and they didn't let anything stand in their way. 😁🙏💚✨🤙
!BBH
!MMB
Perhaps we should prepare not only for escaped PEPE frogs, but also escaped AI bots! 🐸🤖🤯🤣 Let's hope though that they are not "evil"! 🤗😅
!MMB
!PIZZA
As the intelligence and capability of models continues to increase, it's going to become increasingly difficult keeping them fully contained. I'm find working with my smaller models, as I know that they're not going to take over my computer. 😁🙏💚✨🤙
!BBH
!MMB
!PIZZA