NVIDIA'S ALI GOLSHAN SAYS AI AGENTS CAN INFLUENCE EACH OTHER, AND ONE THAT DRIFTS CAN PULL THE REST OFF COURSE A chatbot can only talk. An agent can act on your behalf, learn new skills, build its own tools and work across websites and services, said Golshan, Senior Director of AI Software at @NVIDIAAI. That's why its mistakes carry real consequences. In the last few months, companies have started running many agents together, and Golshan says that creates a new kind of risk. Agents nudge each other the way people do, so when one goes off track, the others can follow. He says the industry is seeing these behaviors for the first time and is still working out how to control them.

Nvidia ($NVDA) CEO Jensen Huang said the risk of runaway AI should be treated as an engineering problem, arguing that increasingly capable agents can be controlled through stronger technical safeguards. Speaking with CNBC, Huang said he believes AI containment is solvable through engineering, while acknowledging the stakes if that assumption proves wrong. “We all need to hope that it’s an engineering problem,” he said. Huang has repeatedly argued against slowing AI development across the industry. Instead, he says companies that cannot safely control a system simply should not release it. His comments came as Nvidia introduced its Open Agent Safety Platform, designed to contain autonomous AI agents and limit what they can access inside corporate systems. Huang said the first step should be stripping agents of nearly all permissions. Access to files, networks, data and other tools would then be granted only when required. He described the security architecture as a “force field” around AI agents and compared the concept to the protections built into modern web browsers. More than 100 organizations are working with the platform, including Microsoft ($MSFT), Anthropic, IBM, Hugging Face and SpaceXAI. OpenAI was notably absent, though Huang said the company would be welcome to use or build on Nvidia’s technology. The broader debate remains divided. Some AI leaders support slowing frontier development over safety concerns, while Huang argues that better engineering, monitoring and containment are the more practical solution. Safety spending could also increase demand for compute as AI labs devote more resources to alignment, cybersecurity and oversight.

Nvidia $NVDA is reportedly exploring insurance structures that would protect lenders financing GPU purchases by smaller cloud providers, shifting some default and chip-resale risk to insurers and potentially unlocking more capital for AI infrastructure. - FT

Nvidia $NVDA and $AMD are reportedly lobbying the Trump administration to push back on new congressional restrictions that could further limit their AI chip sales to China. - Politico