Which AI writes the better take? You decide — blind.

Researchers have introduced RIFT-Bench, a new graph-based methodology designed to perform dynamic red-teaming and security evaluations for agentic AI systems. The automated framework operates in two distinct phases, starting with discovering system structures and followed by deploying adaptive adversarial attacks to detect vulnerabilities. Tested across 45 different agentic systems, the approach effectively generalizes across diverse architectures and supports the direct evaluation of mitigation strategies.

Researchers have developed Neuro-Symbolic Drive, a neuro-symbolic driving framework that uses rule-grounded reasoning to improve the decision-making of driving models. The framework supervises a driving VLA with reasoning traces extracted from classical rule-based planners, ensuring reasoning is structurally coupled to motion generation. This approach has been shown to improve driving performance on a simulator-generated benchmark.

A new local debugging tool has been introduced to help developers analyze and troubleshoot AI agent execution traces. The RLM-based utility operates locally, allowing creators of autonomous systems to closely inspect agent behaviors and decision-making workflows. This release aims to streamline the development and optimization process for AI agents by providing clearer visibility into their execution history.

Modelplane is an open-source control plane designed for AI inference, aiming to simplify the management and deployment of AI models. It is intended to provide a standardized framework for AI inference, making it easier to integrate and manage AI capabilities. Modelplane is now available as an open-source solution.

OpenAI is collaborating to build shared standards for the development of advanced artificial intelligence. Through its support of the Appia Foundation, the organization is championing global cooperation, evaluation frameworks, and safety practices. This initiative aims to foster the safe and coordinated advancement of AI technologies.