Anything you can find with an LLM to break into a locked device/ecosystem, the operators of that ecosystem can find with an LLM too, and fix it. Right?

Only if it's in updateable software. Hardware takes longer to correct.

Hence Sam Altman's latest public comment: "I don't even know a dramatic enough word to use… models can find bugs that are outside of the normal threat model of what we think a hacker can do… hardware-level, insane stuff… I no longer think we're gonna be able to just patch all the bugs."

This is true for things other than LLMs. Outside contributors have been credited with hundreds of security fixes to macOS and iOS that were found with fuzzers and sanitizers. Just off the top of my head CVE-2019-8605. Clearly, Apple has access to AddressSanitizer and libfuzzer, but outsiders still find these things. Not only Apple: Google can find flaws in iOS but outsiders also find flaws in Chrome, so the phenomenon is widespread.