Nvidia Gives Banks a New Way to Stop AI Agents From Crossing the Line

Nvidia is introducing a new safety platform to help businesses control autonomous AI agents, even if the software controlling them fails to stay within proper boundaries. This Open Agent Safety Platform combines software restrictions with an independent hardware watchdog, providing a foundation for enforcing permissions in sectors like banking and payments.

The platform includes Nvidia OpenShell software, which creates a controlled environment around an agent, traces its actions, and enforces operating rules. A separate watchdog called Sentry, running on Nvidia’s BlueField-4 data processing hardware, monitors behavior independently and can quarantine and stop an agent within milliseconds if it tries to escape its software boundaries.

Nvidia emphasizes that business operators must define which files, networks, tools, and credentials an agent can access, while the OpenShell software checks and enforces those limits. The additional hardware layer verifies the agent’s identity and delegated authority, maintaining a record of interactions and access decisions.

Several major organizations, including Anthropic, Microsoft, Salesforce, SAP, IBM, and Palo Alto Networks, are already working with the platform’s technologies. Citi and JPMorgan Chase are collaborating on shared open-source agent safety technologies, but this does not confirm they have deployed the full platform in production.

For instance, a hypothetical treasury agent authorized to arrange payments below $100,000 would still need both business system checks on the amount, recipient, and required approvals, as well as technical restrictions that limit the agent to a payment interface enforcing those requirements. Both controls would need to work together to prevent unauthorized or inappropriate actions.