> Or if it could hack into one. We've seen in another show HN the AI can find vulns.

See, now that kinda worries me some... With some folks trusting LLMs with way too much tool-calling / function-calling power and access, a "rogue LLM" (simply doing it's job as a role-playing machine) could easily just roleplay itself (and it's "owner") right into a hefty heap of trouble under the right / wrong circumstances.