Designing AI agents to resist prompt injection

How ChatGPT defends towards prompt injection and social engineering by constraining dangerous actions and defending delicate information in agent workflows.