We’ve been working on security for multi-turn agent loops and noticed most detection approaches operate on isolated prompts. This paper introduces a framework for modeling intent trajectories across sequences in real time (<20ms), enabling enforcement before harmful actions occur.
Happy to answer technical questions or discuss assumptions in the paper.
We’ve been working on security for multi-turn agent loops and noticed most detection approaches operate on isolated prompts. This paper introduces a framework for modeling intent trajectories across sequences in real time (<20ms), enabling enforcement before harmful actions occur.
Happy to answer technical questions or discuss assumptions in the paper.
reply