NVIDIA AI Updates: July 28, 2026

1. NVIDIA Helps Launch the Open Secure AI Alliance

NVIDIA joined more than 40 inaugural partners to form the Open Secure AI Alliance, an effort to develop and share open source technologies, tools, and techniques for AI safety and cybersecurity. Partners include Adobe, Cisco, Cloudflare, CrowdStrike, Databricks, Dell Technologies, Hugging Face, IBM, Microsoft, Palo Alto Networks, Red Hat, and the Linux Foundation, among others. The group frames its goal as giving defenders open, frontier tools they can control rather than concentrating capability in closed systems. Technical contributions cited include NVIDIA’s NOOA agent framework, HPE’s SPIFFE/SPIRE identity standards, Hugging Face’s Safetensors weight format, and Microsoft’s multi-model scanning work. For security teams, the alliance signals a push toward shared, auditable defensive infrastructure for AI systems. Source

2. NVIDIA Ising Calibration 1.5 Automates Quantum Processor Tuning

NVIDIA released Ising Calibration 1.5, an open source vision language model that interprets diagnostic outputs from quantum processing units and recommends how they should be tuned. The 31-billion-parameter model works zero-shot, scoring roughly 10 percent higher than comparable open models, and improves by 86.68 percent over its predecessor when supplied with in-context examples from related experiments. It is 11.4 percent smaller than prior versions at BF16 precision and ships in an NVFP4-quantized format for single-GPU or DGX Spark deployment. Trained on data spanning superconducting qubits, quantum dots, ions, and neutral atoms, the model is evaluated on the QCalEval benchmark. The approach targets the manual calibration bottleneck that limits scaling of quantum hardware. Source

3. NVIDIA Details Six Agent Harness Capabilities Behind NOOA

NVIDIA published a breakdown of how agent harness design, the architecture surrounding a model, can drive double-digit swings in benchmark results independent of the underlying model. Its NOOA framework represents an agent as a single Python class, with methods as capabilities, fields as state, and docstrings as prompts, and identifies six interface concepts: typed input and output, pass by reference to live objects, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs. NVIDIA reports 82.2 percent accuracy on SWE-bench Verified using roughly half the tokens of comparison systems, 85.1 percent on ARC-AGI-3 for under 20 dollars per game with GPT-5.6-sol, and 86.8 percent on CyberGym L1 vulnerability discovery with no network access. The framework is released open source. Source