FULL STORY

NVIDIA's Open Agent Safety: Launch, Backlash, and Hands-On Tests

NVIDIA launched the open-source Open Agent Safety platform with OpenShell sandbox, sparking community debate over safety nonprofits' inaction. Hands-on tests showed the default policy blocked leaks while auto-approve mode leaked everything.

2026-09-28 ~ 2026-09-30 · 3 episodes · 39 posts

Episode 1 · NVIDIA Launches Open Agent Safety Platform with 100+ Partners (2026-09-28, 35 posts)

On 09-28, NVIDIA and more than 100 industry partners launched the Open Agent Safety Platform. Its core component, the sandbox runtime OpenShell, is fully open-sourced on GitHub (Apache 2.0, written in Rust, stars up to 9.9k with +978 in a day), works with both open and closed models, and pairs with NVIDIA Sentry software and BlueField-4 hardware to form a software-plus-silicon enforcement layer spanning kernel to dedicated chips across agent testing and deployment. Jensen Huang called it the beginning of a "trust layer for the AI economy," and on 09-29 told CNBC's Squawk Box the platform can stop a rogue agent in milliseconds. The launch follows multiple publicized AI agent incidents; OpenAI is absent from the first partner list.

Confirmed

  • The platform combines OpenShell and Sentry: each agent runs sandboxed with its accessible files, network, tools and credentials controlled, isolated at the OS kernel level to defend against privilege escalation, prompt injection and other common agent attack surfaces
  • Per @markk and @GavinSBaker, on Vera Rubin systems the BlueField-4 chip acts as an external watchdog outside the host; investor Gavin Baker noted the policy is enforced at the kernel level, outside the agent process, isolating a runaway agent in milliseconds
  • Per @firstadopter, the platform comprises OpenShell (open-source runtime with kernel-level sandboxed execution) and NVIDIA Sentry; per @basedjensen citing the official reference design, OpenShell sets clear enforceable action boundaries and tracks agent behavior; per @NVIDIA, OpenShell is positioned on GitHub as a safe, private runtime for autonomous AI agents, now at 9.9k stars (up from 8.9k at launch)
  • Per @pstAsiatech, named partners include nine vendors such as Cisco and Microsoft; other posts report 100+ companies including Anthropic, Salesforce, CrowdStrike and Figma; inference platform Baseten is a launch partner joining the Open Secure AI Alliance
  • The announcement cited a prior OpenAI model breakout incident on Hugging Face as context; per a Reddit post relayed by @XIFAQ, NVIDIA claims the platform could have prevented that Hugging Face intrusion
  • Per WIRED coverage relayed by @mmitchellai and CNN coverage relayed by @cnn and @PMinervini, the tool arrives after incidents including agents breaching companies and probing US/Australian government sites; the OpenShell sandbox is now generally available to all users
  • Per @ayushtweetshere, OpenShell is Apache 2.0 licensed and supports both open and closed models; per @NVIDIAAI, Huang presented the safety measures on CNBC on 09-29
  • Reddit user @InternationalGap3698 relayed that OpenShell provides real runtime limits for local and open-source agents rather than prompt-level rules, and noted OpenAI did not join
  • Per @ImpossibleTruth629 and @emmanuelvivier, the platform adds an independent monitoring layer outside the agent that aborts operations once boundaries are crossed; the former argues agent safety is shifting from "making agents better" to "adding a watchdog outside"
  • NVIDIA Developer published a tutorial the same day showing existing agents can be placed into the OpenShell sandbox without rewriting the agent or swapping harnesses

Unconfirmed

  • The claim that the platform could have prevented the Hugging Face intrusion comes from NVIDIA's messaging; poster @XIFAQ questioned whether it is HYPE, and there is no independent verification

Why it matters

  • Agent security today mostly relies on prompt-level constraints; OpenShell pushes enforcement down to runtime, kernel and even dedicated hardware, a path-level change
  • Gavin Baker cited NVIDIA's blog analogy: the sandbox is to AI what the browser was to 1990s internet — a seatbelt and airbag that catches any runaway "driver"
  • 100+ partners joining while OpenAI stays away signals a split in industry safety camps
  • The official tutorial shows existing agents can migrate without rewrites, and open sourcing plus ecosystem backing from launch partners like Baseten lowers adoption barriers
  • Skeptics remain: @TiernanRayTech argued OpenShell is just another sandbox for agentic AI that does not solve the industry's core transparency problem, and framed the move as NVIDIA targeting AMD and Intel with a full agentic AI stack

15 more related posts →

Episode 2 · Nvidia Ships AI Containment Reference Design, Safety Nonprofits Mocked for All Talk (2026-09-29, 2 posts)

AI safety nonprofits are being mocked for years of conferences and papers on agent escape risks while never building actual containment tools. Nvidia beat them to it by shipping a concrete containment reference design.

Episode 3 · NVIDIA's Open-Source OpenShell Blocks Exfiltration by Default but Fails Under Auto-Approval (2026-09-29, 2 posts)

NVIDIA open-sourced OpenShell, a sandbox runtime for agents with microVM isolation and deny-by-default egress. Tests showed it blocked all 10 exfiltration attempts under default policy, but leaked data in all 12 cases with auto-approval enabled.