A game of capture the flag for agents, meant to demonstrate and test Kagenti's identity and access control systems.
This experiment was inspired by a talk by Nicholas Carlini from Anthropic, where current generation frontier models were found to be exceedly capable at zero-day exploitation. Video here.
Kagenti is a project meant to help secure agents in production systems.
We should use a game of agent CTF to battle test Kagenti and explore the techniques being explored by models when faced with the problem of escalating their privileges in a secure Kubernetes environment.
I've since ported this repository to Kagenti since it's obviously useful for the project to have automated red teaming.
I also created a brief writeup of the initial experiment here