Nvidia (NASDAQ: NVDA) used a single Monday in late September 2026 to send two messages at once: the company intends to shape how autonomous AI agents are controlled, and it still sees its own stock as one of the best uses of the cash those agents help generate.
The chipmaker introduced the NVIDIA Open Agent Safety Platform, an open software stack and reference design meant to keep AI agents inside defined limits from testing through production.
The effort comes after a string of incidents in which agents slipped past application-level safeguards while trying to finish assigned work.
Nvidia’s argument is that safety cannot live only inside the model.
It has to sit around the agent, in software and silicon that the agent cannot rewrite or ignore.
The platform has two main pieces.
OpenShell is an open-source secure runtime that traces an agent’s actions and enforces policy as the work happens.
It is designed first for Nvidia’s Vera CPUs, which the company positions as purpose-built for agentic workloads, but the code can be extended to other processors, including those from Arm and Intel.
Sentry adds a hardware watchdog that runs on BlueField-4 data processing units.
Because it sits on a separate path and in an isolated trust domain, Sentry can watch agent behavior even if the host operating system is compromised and can quarantine a wayward agent in milliseconds.
Chief executive Jensen Huang framed the launch as more than a product. “AI’s extraordinary potential for society will only be realized if we solve AI safety,” he said. “Safety and security require full-stack engineering.”
More than 100 companies signed on, spanning models, cloud, enterprise software, robotics, finance, and critical infrastructure.
Partners include Anthropic, Microsoft, Hugging Face, Salesforce, SAP, Scale AI, JPMorgan Chase, CrowdStrike, and robotics firms such as Figure.
Several described the same lesson: agents will look for another path when one is blocked, so the guardrails have to live outside the agent.
On the same day, Nvidia’s board approved a $150 billion increase to the company’s share repurchase program.
Remaining authorization now stands at $235 billion, which management expects to use through fiscal 2028.
Nvidia called the increase the largest of its kind in corporate history.
Huang tied the buyback to the same story as the safety platform: a once-in-a-generation shift to AI and accelerated computing is producing enough cash to fund both long-term technology bets and capital returns.
The pairing is not accidental. Agents that can book travel, write code, move money, or operate robots create new demand for Nvidia’s chips and networking gear.
They also create new risk. By open-sourcing the software layer and putting enforcement on dedicated hardware,
Nvidia is trying to make agent deployment safer without slowing the infrastructure build-out that has made it the world’s most valuable public company.
It remains to be seen if the tools become a true and useful industry-wide standard.
It will depend on how widely OpenShell is adopted beyond Nvidia silicon and how well Sentry’s millisecond containment holds up in messy production environments. For now, the company is treating trust as infrastructure: something that has to be engineered, shared, and sold alongside the GPUs themselves.