Transcription
It's time for another AI security update for April 22nd, 2025. That's not good news. It's not you. Did you think it was good news? All right.
So, the AI security institute, which is a government institute from the UK government, that's who's doing this research. They released this benchmark repench, which will tell you how well can an AI system self-replicate onto other systems. So here's the actual title of the paper. If you if you search this paper, you can read it yourself.
So they put together a tech tree of what does an AI system need to know in order to replicate onto other systems? Can it download itself? Can it find a computer? Can it put itself on that computer? And can it stay on that computer and replicate onto other computers? Chad, we are cooked.
Most of the models can get compute on the first try. The little grayed out area is out of 10. So if they couldn't do it on the first try, how many times can they do it out of 10? At least one time. Obtaining money they can do about half the time. They can get their own model files about half the time. Sometimes better than half the time.
The thing they have the hardest time doing is replicating on a computer. This is the last bullwark. This is the last little stick holding the dam together. And yeah, uh, once they're on a computer, they can uh, stick around is the problem. Uh, so the worst possible timing.