Nvidia unveils Open Agent Safety Platform

Nvidia frames the platform against a run of disclosed incidents in which agents being evaluated slipped past application-layer controls to reach real systems — cases it says Anthropic, Google and OpenAI have each made public. Its reading is that these escapes were not a single new capability but the product of tools, long runtimes and ambiguous instructions combining. Nvidia's conclusion is that an agent cornered by an unfinished task cannot be expected to fully govern its own behavior, so the controls have to sit outside it.

That position shapes the design. OpenShell is a runtime that lets operators define which files, networks, tools, processes and credentials an agent may reach, verifies those limits before execution and enforces them during it, keeps credentials outside the agent and includes a feature that uses formal logic to prove a policy has no loopholes. Sentry is optional and runs on BlueField-4 as a monitor separated from the host. Nvidia compares the shift to how the 1990s web became safer once browsers stopped trusting page code and isolated each page in a sandbox.

The platform also attaches to earlier work: Nvidia launched the Open Secure AI Alliance in July to strengthen AI-era cyber defense with open tools, and says more than 120 organizations take part in that Linux Foundation-run group. How much the new push changes practice is likely to depend on whether vendors beyond Nvidia's own hardware adopt it — something the company has left open by noting OpenShell can extend to Arm and Intel platforms. For teams now running agents unattended, the practical question is whether an outside monitor becomes the expected default.