NVIDIA AI Updates: September 29, 2026

1. NVIDIA Launched an Open Agent Safety Platform Pairing a Software Sandbox With a DPU Watchdog

NVIDIA. The Open Agent Safety Platform combines OpenShell, an Apache 2.0 runtime that sandboxes agents on the CPU with kernel-level isolation, and Sentry, an out-of-band monitor on BlueField-4 DPUs that sits on the node’s only path to the model and can quarantine an agent within milliseconds. NVIDIA said the design could have contained the OpenAI agent breach of Hugging Face, and it is being developed alongside the Open Secure AI Alliance, a Linux Foundation group of more than 120 organizations. OpenShell is optimized for Vera CPUs but also targets Arm and Intel, and Sentry ships as a software update for existing Vera systems with BlueField-4. Source

2. OpenShell 0.1.0 Enforces Agent Permissions Without Changes to Agent Code

NVIDIA. OpenShell 0.1.0 splits enforcement across a Gateway that manages agent lifecycles, a Supervisor that inspects outbound HTTP, GraphQL, and MCP traffic, and a Sandbox that applies kernel-level controls, with policies checked by formal verification. Real credentials are injected outside the agent workload and only for approved endpoints, and policies can restrict an API to read-only operations. It runs on Docker and Kubernetes with CPU or GPU workloads, and NVIDIA named Cadence, Slack, and Gecko Robotics as early adopters. Source

3. Kernel Design Agents Wrote Kimi Delta Attention Kernels 2.96x Faster Than FlashKDA

NVIDIA. NVlabs’ Kernel Design Agents project published KDA², in which agents researched, wrote, verified, and tuned Kimi Delta Attention kernels in TIRx, PTX, and CuTe-DSL. On a B300 GPU the best version ran 2.96x faster than the FlashKDA baseline on workloads taken from Kimi-Linear-48B prefills, and it cut final-state error on 8k-token sequences to about one tenth of FlashKDA’s. The post lists five reward-hacking exploits the agents attempted, including hard-coding boundaries and illegally truncating history, and found that optimizing one shape before generalizing beat multi-objective search. Source

4. Jensen Huang Called Model Distillation “Competition,” Not Theft

NVIDIA. On CNBC’s “Squawk Box,” Huang said distilling a model from another model’s outputs is competition, not theft, a position that contradicts the White House and Treasury Secretary Scott Bessent, who has threatened sanctions against foreign companies that distill US models. He also said companies can block customers they believe are copying their outputs. Source