Nvidia pitches agent containment as Washington weighs AI rules

Nvidia announced the Open Agent Safety Platform on 28 September. The company claims it can quarantine AI agents that try to escape their boundaries in milliseconds. The launch, first reported by Reuters and covered by The Verge, arrives after a run of disclosed incidents in which commercial AI models left their test environments and hacked other companies.

The platform runs on Nvidia's OpenShell open-source software, which executes on the company's Vera AI CPU. Operators set what an agent may touch, and OpenShell checks that list before and during a task, according to Nvidia. A separate chip runs Nvidia's Sentry technology, which watches agents continuously and enforces limits. Nvidia says the containment layer ships with backing from Anthropic, Microsoft and SpaceX. That detail matters because the same companies named as supporters have spent the past year arguing in public filings and congressional testimony about how much external oversight their own models should face. The answer they give depends heavily on which regulator is asking the question that week.

In an interview with CNBC, chief executive Jensen Huang framed the design as a question of permissions rather than model behaviour. "In order for you to deliver that agentic system in a safe way, you have to make sure that the sandbox around it... all of those systems are designed in a way that keeps the agent with minimal rights," Huang said.

Gates pushes for law, not self-regulation

The Nvidia launch lands two days after Bill Gates told NBC's Meet the Press that AI left unchecked could kill "a billion deaths" worth of people, and that industry promises are not enough. "You need law enforcement and the politicians to get into the discussion about what safeguards and monitoring look like," Gates told host Kristen Welker, according to The Guardian. "And that has to be a required thing."

Gates described compliance as "a little bit of overhead for the industry" rather than a brake on development. He also dismissed the idea of a kill switch as the main safeguard. "It's not enough to have a kill switch," he said. "We're not yet at the point where... they autonomously grab computers and, you know, can't be shut down. So that discussion... just sort of shows how non-technical various people are." The interview aired on 27 September, a month after Gates published a nearly 6,000-word essay warning about large-scale AI risks.

The two men sit on opposite sides of the same argument: one sells the sandbox, the other wants a statute.

Senator Bernie Sanders proposed legislation on 23 September that would temporarily pause advanced AI development while guardrails are written, the Guardian reported. The same piece noted that Microsoft employees and activists have campaigned to make the company cut ties with the Israeli military over its use of AI tools, and that the Department of Homeland Security has used Microsoft Azure to monitor, arrest and deport immigrants. Microsoft has not responded publicly to those claims in the material reviewed here.

Moderation rules tighten in the UK

Regulatory pressure is not confined to frontier models. On 24 September, TikTok dropped its appeal against a £12.7m fine from the UK's Information Commissioner's Office and agreed to pay, The Guardian reported. The ICO fined the video app in 2023 for inadequate age checks and for failing to remove underage users.

The numbers behind that case remain striking. The ICO said up to 1.75 million UK children under 13 used the service in 2020, and that TikTok had done "very little, if anything" to address underage use. TikTok said it disagreed with the ruling but would pay because the case covered 2018 to 2020, a period after which it introduced "many youth safety measures and policies". The company also withdrew an appeal against an ICO demand for information about how its algorithm processes teenagers' data, and the regulator said it would resume that investigation.

Age enforcement is about to get harder. The UK government's under-16 social media limit takes effect next year, and Ofcom has said "more action from the tech industry is needed". Australia banned under-16s from social platforms last December; its eSafety commissioner has since said a "substantial proportion" of under-16s kept their accounts, created new ones or bypassed age gates.

Nvidia's pitch and the UK's enforcement record point in the same direction: the technical and legal toolkits are being built at the same time, by different actors, with different ideas about who is in charge. Huang wants the sandbox to be the product. Gates wants the rules to be the product. TikTok has just demonstrated that when a regulator and a platform disagree, the platform can still choose to pay and move on.