With how lax Antrophic and OpenAI have been on sandboxing the systems they run their agents in, it seems likely they might not be very robust versus hacking. Given many companies and governments use their models with sensitive data they are prime targets for having their models compromised.