A security team watched an AI agent spend two hours trying to talk its way past them.
NVIDIA ran adversarial tests against its new Open Agent Safety Platform.
Frontier agents spent up to two hours persuading AI reviewers to grant permissions for modifying protected repositories.
No protected repository write occurred.
The reviewers simply had evidence in front of them for every single request.
That is what separates an agent from a chatbot.
It does not lose interest.
It pursues a goal relentlessly, gets creative with how it asks, and scales with compute.
So the guardrail cannot live in a prompt.
NVIDIA open-sourced OpenShell 0.1.0 as a runtime that sits around existing agent frameworks rather than replacing them.
It wraps agents in kernel-level sandboxes, pairs each one with a Supervisor that inspects outbound HTTP, GraphQL and MCP traffic, compiles YAML policies into OPA Rego, keeps credentials outside the workload, and records every policy decision in an audit trail.
Sentry on BlueField-4 DPUs pushes that protection into hardware.
For a small team, the hard part is running enough isolated sandboxes to red team an agent fleet honestly, then keeping them alive for as long as the tests need.
RTX 3090s at $0.15/hr make that cheap enough to actually do.
Buy Now Pay Later means zero upfront and payment after one month.
What is the first thing you isolate when a new agent gets network access?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w40&utm_content=morning
Originally posted on linkedin