> if a malicious actor can weaponize an agent to do their bidding

In my experience, human employees are much more vulnerable to this particular weakness than frontier agents (i.e. phishing attacks).