Computers don't "know" anything. Computers are gadgets for executing sequences of arithmetic operations & can be formalized w/ any number of abstractions like Turing machines, lambda calculi, context aware grammars, cellular automatons, stack machines, Diophantine equations, etc. All of these models are equivalent in terms of computational capabilities & none of them can be said to "know" anything.

The general confusion about AI stems from the fact that people are unwilling to learn the bare minimum amount of computability theory to make sense of abstract algorithms & their physical instantiations on computers. There are no ghosts in the machine, there is no spooky self-awareness that arises from a large enough number when it is interpreted as the code for some algorithm on some machine. What's happening here is that there are common statistical patterns in benchmarks that Anthropic has inadvertently encoded into their neural networks. They didn't have to train on benchmarks but they did & now the regular statistical patterns of benchmarks are biasing whatever computation they are doing to certify their software for "safety" (or whatever is the latest euphemism for not telling people how to make nukes).