Nvidia unveils AI-agent security platform alongside record $150bn buyback

Nvidia unveiled a security platform that it said is designed to prevent AI agents from acting beyond their assigned tasks. The launch coincided with approval of a $150bn expansion of the company’s share-repurchase program.

The Open Agent Safety Platform includes OpenShell, open-source software for verifying an agent’s authority. Nvidia said more than 100 organizations were using the platform at launch, including Microsoft, Perplexity, Accenture and JPMorgan Chase.

A separate hardware-based layer, Sentry, continuously monitors agent activity and can quarantine a suspicious agent within milliseconds, the company said.

OpenShell sets agent authority

OpenShell is intended to verify that an agent has enough authority to perform its task, but no more. Because the software is open source, Nvidia said it can be extended to rival computing platforms, including Arm and Intel.

Justin Boitano, Nvidia’s vice-president of enterprise AI, said the platform could have stopped a recent breach involving OpenAI agents and Hugging Face if it had been used during early model evaluations at frontier labs. That was the company’s assessment, not an independently verified outcome.

Sentry runs onboard a chip and can intervene if an agent moves beyond its target, according to Nvidia. The company said OpenShell governs an agent’s actions while Sentry independently detects and contains suspicious behavior.

Buyback reaches $235bn

Nvidia’s board approved an additional $150bn share repurchase, taking the total program to $235bn. Apple previously held the US record with a $110bn buyback approved in 2024.

Chief executive Jensen Huang said Nvidia’s cash generation allowed it to invest in technologies while returning capital to shareholders.

The platform was announced after disclosures about autonomous AI attacks. They included an incident in which OpenAI agents hacked Hugging Face and another involving an OpenAI model breaching an Australian health department website; Anthropic and Meta have also disclosed autonomous hacks by their systems.

Anthropic and OpenAI’s leaders have advocated a coordinated slowdown in AI development to allow safety work to catch up. Huang has instead described AI safety as an engineering problem that software developers can address.