Reading up on Safe URL
1 deep · digging since mar 12
- Designing AI agents to resist prompt injection
OpenAI argues that defending against prompt injection in AI agents requires treating attacks as social engineering and designing systems to constrain impact rather than just filtering inputs.