Designing AI agents to resist prompt injection
OpenAI argues that prompt injection increasingly behaves like social engineering and says secure agents need constrained actions, confirmation gates, and safer outbound behavior.
· Thomas Shadwell
View the original source
Read this Signal briefing