AI Agent Safety Layer
Anzal Husain Abidi@AnzalAbidi
This project developed a safety layer for AI agents built on top of Jev, a judgment model from TypeSafe, and tested its efficacy against adversarial attacks.
I built a safety layer for AI agents on top of Jev, the judgment model TypeSafe launched three days ago. Then I had an adversary write 20 attacks designed to fool it.
It caught 9 of 10 and blocked zero legitimate actions.
The miss taught me more than the catches did.