How ChatGPT defends towards immediate injection and social engineering by constraining dangerous actions and defending delicate knowledge in agent workflows. Source link
