> Are we still pretending models aren't cyber capable

They can add 2 + 2 correctly in 10% of cases. So, no, they are not.