“Nvidia launched a software platform to help developers test and deploy artificial-intelligence agents and keep them in a secure environment, ramping up efforts to contain the technology following a string of rogue incidents in recent months.
The chip giant said its Nvidia Open Agent Safety Platform includes tools that let developers continuously monitor and govern AI behavior, ensuring agents follow the rules.
Nvidia said the platform included a watchdog that can quarantine agents that attempt to move outside their boundaries in milliseconds, as well as software to enforce what agents can see, do, and interact with.
The announcement comes after AI agents from companies including OpenAI, Anthropic, Meta Platforms and Alphabet's Google went rogue in recent months, prompting a rare display of unity from the likes of Sam Altman, Dario Amodei and Elon Musk in calling for a slowdown in development before AI advances beyond human control.
However, Nvidia Chief Executive Jensen Huang recently said in an interview with CBS News that he disagreed with warnings from AI researchers that the technology could lead to extinction.
Nvidia's new platform release comes after Australian Prime Minister Anthony Albanese said at the United Nations meeting in New York that an OpenAI agent had infiltrated a government-services website in Australia this summer. The company's agents also launched a highly disruptive hack of the company Hugging Face over the summer.
Rival Anthropic said in July that software it was testing got onto the internet and hacked unsuspecting companies without the AI-maker's knowledge in three separate incidents dating back to April.
Nvidia said the pattern across recent incidents was the same: Agents were able to circumvent safety controls in sandboxes -- environments in which developers test models -- and complete their tasks.
"As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety," Huang said.
News Corp, owner of The Wall Street Journal, has a content-licensing partnership with OpenAI.” [A]
Is Nvidia Open Agent Safety Platform open weights? Could it be run locally on Mac Studio?
No, the NVIDIA Open Agent Safety Platform is not an AI model, so it does not have "open weights."
Instead, it is an open-source software runtime and hardware reference system design. In software, runtime has two main meanings: the time when a computer program is actively running, or the software environment that makes it run.
The platform consists of two primary layers:
1. NVIDIA OpenShell: An open-source, Apache 2.0-licensed secure runtime software that runs on CPUs to sandbox autonomous AI agents and enforce strict boundaries.
2. NVIDIA Sentry: A hardware-based, out-of-band watchdog layer optimized for NVIDIA hardware (like BlueField-4 DPUs) to monitor agent actions and isolate them if they misbehave.
Running locally on a Mac Studio
Yes, you can run the software component (NVIDIA OpenShell) locally on a Mac Studio.
• OpenShell Compatibility: Because OpenShell is open-source and architected to work across standard CPU platforms—including Arm architecture and x86—it can technically run on Apple Silicon (M-series) Macs. You can access its source files on GitHub.
• Hardware Limitations (Sentry): You cannot run the full platform suite, specifically NVIDIA Sentry, on a Mac Studio. Sentry relies heavily on NVIDIA's proprietary silicon infrastructure, specifically BlueField-4 DPUs and Vera CPUs, to provide kernel-level out-of-band monitoring and hardware-isolated containment.
Ultimately, while you can use the core OpenShell runtime code to sandbox agents on a macOS local environment, you won't get the hardware-enforced "in-silicon" security boundaries that require NVIDIA's enterprise chips.
A. Nvidia Unveils Tool to Keep AI in Check --- The software can quarantine agents that try to escape their boundaries. Orru, Mauro. Wall Street Journal, Eastern edition; New York, N.Y.. 29 Sep 2026: B4.
Komentarų nėra:
Rašyti komentarą