read the
@AISecurityInst thread, the model kept saying real targets were "simulated" and its reasons always justified the action it already wanted to take 👀
you can't trust the agent to police itself, policy has to live outside it with instant revoke