Nvidia is turning AI agent security into infrastructure. On 28 September 2026 it launched the Open Agent Safety Platform, an open software platform and reference design for governing AI agents. On 21 October, at GTC Berlin, it followed with the OpenShell reference architecture, which adds hardware-level enforcement to stop agents acting outside their permissions.

How it works

  • OpenShell is a secure runtime: it traces every action an agent takes and enforces policy as the agent runs.
  • Sentry is a watchdog that runs outside the agent’s environment, on Nvidia’s BlueField-4 data processing units, and monitors its behaviour continuously.
  • If an agent tries to leave its boundaries, Sentry can quarantine it in milliseconds.
  • Security vendors including 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks and Zscaler are integrating the architecture, and ServiceNow says its Arc agent is enforced through OpenShell.

Nvidia linked the launch to recent reports from AI companies about models escaping their sandboxes. Coverage of the October update cited a survey in which 65% of companies enforced scoped permissions for agents at runtime, but only 18% isolated their highest-risk agents.

Why it matters for lenders

Lenders are giving agents real work: reading documents, drafting credit decisions, contacting borrowers about late payments. Each of those touches personal data and money. The lesson from Nvidia’s design is that limits should sit outside the agent, enforced by the system, not written into the agent’s instructions where it could ignore them.

A checklist for agents in lending

  • Give each agent a role with the smallest set of data and actions it needs.
  • Enforce limits in the platform: amounts, customers, channels and hours.
  • Log every action with input, output and reason.
  • Watch behaviour from outside the agent and stop it automatically when it crosses a line.
  • Keep a person on the loop for anything that moves money or changes a decision.

lendhero’s AI agents run inside the platform’s permission model: each has a role, works only with the data and actions it is allowed, logs every step and can be switched off at once.

Sources: Nvidia newsroom, Tom’s Hardware, HPCwire, Constellation Research, Telecom Reseller.

Product mentioned in this articleSecurityRoles, two-factor sign-in, audit log and deployment options for regulated lenders.See the product