I do hope for a future where LLMs are so good at patching zero-days that the NSA and all other nation-state 'cyberwarfare' departments (ie software privateers) lose most their capability.
On the other hand, nation-states will be able to run massive GPU clusters full of LLMs dedicated to cyberwarfare. They certainly have the resources to do it.
Imagine the Huggingface incident, but with an agent swarm that's deliberately malicious instead of just misaligned.
Seems just as likely that LMMs will have the opposite effect.
Zero-days are the least of our worries in the late 2020s and early 2030s