the Muse agent is deliberately separated from credentials, network permissions, and security enforcement, so compromising or manipulating the model should not automatically give an attacker access to everything. https://t.co/jr8By10gIZ