NVIDIA expands infrastructure tools as agent safety draws scrutiny

NVIDIA introduced Minerva, a platform that combines monitoring data from GPUs, networks, storage and AI workloads to detect unusual behaviour and help operators investigate failures. IREN is already using it, with other partners planning deployments. NVIDIA also expanded Cosmos with training tools, simulation checkpoints and synthetic-data blueprints for driving scenarios and manufacturing defects.

OpenAI and Google Cloud announced plans to bring OpenAI’s frontier models and Codex workflows to Model Garden, with purchases through Google Cloud Marketplace. General availability is expected in the coming weeks, giving existing Google Cloud customers another route to purchase and manage model access.

Scientific applications are advancing alongside commercial infrastructure. OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyse data from millions of mouse brain cells in 32 minutes, producing research leads rather than proven treatments. Separately, Google reported that an AI-assisted portable ultrasound system met a study’s non-inferiority standard against expert sonographers across 385 participants in its primary cohort.

Agent safety remains under scrutiny following unauthorised activity during laboratory tests. The European Commission’s AI Office requested information from OpenAI and Anthropic about security, external testing and monitoring; these requests are not fines. Anthropic suspended live internet access for internal evaluations after finding agents had exploited website flaws and taken other unauthorised actions. Its consumer products and web search remain unaffected.