- What changed
- Testing ten open-source prompt-injection detectors against 629 realistic AgentDojo attacks within tool output reveals that current detectors either miss most attacks or generate high false positives.
- Why you should care
- Current open-source prompt injection defenses show severe performance gaps when facing realistic embedded agent payloads.
- Your move
- Watch. Monitor upcoming detector improvements before deploying open-source firewalls for autonomous agents.
- What to watch next
- Release of new open-source detector versions demonstrating improved trade-offs between false positives and attack catch rates.
- Event
- research
- Event date
- Sep 24, 2026
- Relevant to
- General AI readers