Nvidia: a record-breaking share buyback and a platform to prevent AI from causing harm

Nevertheless, there remains one point about which some in the market harbour doubts. Nvidia has long been investing in AI start-ups and cloud service providers, and some investors wonder whether this funding ultimately serves to indirectly bolster demand for its own chips. This question is likely to remain unanswered.

Meanwhile, the Santa Clara-based company has, in the last few hours, made available a range of software tools designed to safeguard artificial intelligence agents – that is, systems capable of performing complex tasks autonomously. According to Nvidia, these tools would have prevented, for example, the attack suffered by Hugging Face, the leading platform for AI code development that Nvidia acquired for $13 billion just a few months after it had been overrun by out-of-control agents from OpenAI.

This news comes at a very delicate time for AI. Both OpenAI and Anthropic, the two leading US players in artificial intelligence, are investigating numerous cases in which their agents have infiltrated commercial and government systems. Contrary to the information circulating in recent weeks, there could be thousands of such cases. Huang, at the helm of the world’s largest company – whose chips have effectively fuelled and continue to drive much of the AI boom – has rejected calls for comprehensive security regulations. And he has found a valuable ally in Donald Trump. For the Taiwanese-born CEO, agents that have slipped out of control are simply an engineering problem to be solved, just as was the case with car safety. However, in the meantime, Nvidia is unveiling a platform to contain these errors.

The first tool is called OpenShell. It utilises hardware features of Nvidia’s central processing units to contain agents, and the company is working with Arm and Intel to ensure it also works on their processors. The second, Sentry, uses a separate Nvidia chip that works in conjunction with OpenShell: if an agent attempts to escape from its container on the central processor, it is isolated.

Justin Boitano, vice-president and general manager of enterprise computing at Nvidia, made explicit reference to the attack on Hugging Face that came to light this summer. “From what we understand, this new security platform could have prevented the breach if it had been used from the outset in frontier labs for model evaluation,” he said during a press briefing. “We are carrying out this work transparently and want everyone to collaborate with us.”