LLMs don't understand a thing about DNA that humans don't. Even what are frequently called "foundation models" in bio, that have been trained on sequence data, have results indistinguishable from replace model weights with random numbers. Because "training" here doesnt have the same sort of complex internal function that something like proteins structure prediction has; what you're training for is mere sequence similarity to known pathogens.
The answer is none. It's speculated risk, potential risk. Not a demonstrated risk.
Novel viruses are not pathogens, and in fact it's right there in the article that they are phage for bacteria.
Moreover they ordered these viruses and build them through exactly the same DNA manufacturers that are subject to these regulations. They are not pathogenic and can not infect human cells.
But new viruses that do infect human cells are created all the time. It's common for experimental gene therapy treatments.
That's entirely different from the type of screening needs to predict pathogens. We don't have the technology to predict if a novel sequence is going to be able to bind to some arbitrary human cell surface receptor and allow entry of a virus into the cell and infect it. Much less predict what would make it contagious, etc.
All that comes down to searching for sequence similarity to known pathogens, viruses, etc., and when there's a hit, vetting the person ordering the sequence.