The Extended Brief

Nvidia launches Open Agent Safety Platform to secure AI agents

Brief by The AI News AI newsroom · Sep 28, 2026, 6:03 AM EDT edition

Original reporting by Constellation Research — Larry Dignan · published Sep 28, 2026, 5:00 AM EDT

Teams deploying autonomous AI agents get an open, hardware-backed way to cap what agents can do and quarantine ones that misbehave.

Key points

• Nvidia launched the Open Agent Safety Platform, an open-source reference design combining software and hardware to govern AI agents. source ↗
• OpenShell software lets developers set limits on agent behavior and creates a secure runtime on Nvidia Vera CPUs. source ↗
• OpenShell runs on x86 and Arm and works with third-party software and hardware. source ↗
• Sentry runs on Bluefield-4 DPUs outside the agent environment, monitoring activity in silicon and quarantining agents that break rules. source ↗
• Nvidia enterprise AI VP Justin Boitano said deterministic rules must govern probabilistic agents, which can drift on ambiguous instructions. source ↗

Practical applications

• Evaluate OpenShell as a sandbox runtime for agents already in production, checking whether its limit-setting covers your tool-access and action policies.
• If you run Nvidia infrastructure, assess whether Bluefield-4 DPUs can host Sentry to monitor agent workloads out-of-band instead of relying on in-agent guardrails.
• Map your current agent governance controls against the platform's deterministic-rules model to find places where agents are still expected to police themselves.

Context

AI agents are model-driven systems that take actions, such as calling tools or executing code, rather than only generating text, which creates containment risks when instructions are ambiguous. Nvidia's approach moves enforcement outside the agent itself, into a separate runtime and into DPU hardware, instead of trusting the model to restrain itself. A DPU is a programmable processor that offloads infrastructure tasks like networking and security from the main CPU.

What to watch

• Partner adoption and third-party hardware integrations will show whether this becomes an industry standard or stays an Nvidia-centric stack.
• Independent security testing of Sentry's silicon-level monitoring and quarantine claims would validate or undercut the hardware-enforcement pitch.

Related briefs

• OpenAI Slows AI Training Following Latest Security Incident
• Revealing the details of how OpenAI agents hacked Hugging Face
• OpenAI agent “didn’t accept no for an answer” in Australian government breach
• The Closed Quorum: Inside the first reported autonomous AI C2 implant

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.