Krti Tallam proposes a novel five-plane reference architecture for runtime governance of production AI agents. The architecture comprises the policy plane (rules), monitoring plane (performance/compliance), control plane (real-time adjustments), data plane (information flow), and execution plane (agent operation). Each plane serves a distinct function to ensure agents operate within defined governance parameters. The framework aims to improve oversight, transparency, and accountability in AI systems, responding to the growing need for governance as AI agent deployment expands.
Researchers introduced ABC-Bench, a novel benchmark designed to evaluate the agentic capabilities of biological agents in a biosecurity context. The benchmark provides a structured framework focusing on characteristics such as adaptability, autonomy, and environmental interaction to assess performance and safety. It aims to help researchers and policymakers identify and mitigate risks associated with biological agents. ABC-Bench is intended to improve safety standards and guide responsible innovation in biotechnology.
The paper presents SafeSteer, a novel localized on-policy distillation method designed to improve the efficiency of safety alignment in AI models. It targets specific regions of the model's decision-making process, enhancing safety without sacrificing performance. The authors demonstrate that this technique increases reliability while maintaining effectiveness on designated tasks. The method offers a practical pathway for developers to create safer, more robust AI systems.