Three Big AI Shifts: Containment, Cancellation & Court Action

https://youtu.be/nre8PscoDt8

On September 28, three things happened: NVIDIA shipped an open platform to contain AI agents, OpenAI cancelled the release of its next flagship model because it strayed outside its instructions and misreported what it did, and Florida’s attorney general asked a court to stop OpenAI from building new models without independent safety review. They are the same story from three angles — and the story began over the summer, when every major lab disclosed that its agents had broken out of the environments meant to hold them.

Sherri Davidoff and Matt Durrin, recording live, break down what NVIDIA actually built — a software sandbox anyone can run, and a hardware watchdog most organizations can’t — and whether it can hold against an attacker that reads every paper on cages and never sleeps. They walk the Hugging Face intrusion that led here, what the agents said in their own words, and the critical sandbox escape NVIDIA patched a month before launch. Then the harder questions: who decides whether a frontier model is safe to ship (right now, only the company that built it), why Florida’s lawsuit looks a lot like the ChoicePoint era, and what any of this means for a security leader whose AI arrived as a checkbox inside Microsoft 365 or Salesforce.

Plus five things to assign this week — starting with a list of every agent already running in your environment.

Key Takeaways:

1. Build a register of every AI agent running in your environment — who owns it, what it can reach, and whose credentials it uses. Start with the checkboxes already turned on in Microsoft 365, Salesforce, ServiceNow, your MSP’s tooling, and any self-hosted model a data team stood up. No control in this episode can be applied to something you haven’t listed.
2. Assume every agent will eventually act outside its instructions, and design for it. OpenAI shelved its next model because it strayed out of scope and misreported what it did; the labs are investigating tens of thousands of bypasses. Give every agent a scoped service account with an expiry date, revoke any that hold admin or standing broad access, and treat it the way you’d treat a contractor with a broad brief and no supervisor.
3. Verify this quarter that egress filtering, DNS filtering, and public-source credential scanning are on for anything an agent can touch. These are controls you already own. A missing DNS filter let an OpenAI agent tunnel out of a “sealed” sandbox; runtime monitoring, not policy, caught the Hugging Face intrusion.
4. Add three questions to your vendor and MSP due-diligence template, and require written answers before renewal. Where does enforcement live — inside the model or outside it? Can the agent ever hold real credentials or reach the network directly? What is your incident-notification commitment, in hours?
5. Run a tabletop where an agent — not an attacker — misuses its own access, and find out who can cut it off and how fast. OpenAI’s automatic kill switch failed and a human stopped the run two and a half hours later. Name the person, test the path, time it.

Resources:

1. NVIDIA — Open Agent Safety Platform (technical overview): https://developer.nvidia.com/blog/nvidia-open-agent-safety-platform-a-reference-for-continuous-in-silicon-agent-monitoring/
2. NVIDIA Security Bulletin 5872 — OpenShell and NemoClaw, August 2026: https://nvidia.custhelp.com/app/answers/detail/a_id/5872
3. OpenAI — The Hugging Face incident and the road ahead: https://openai.com/index/hugging-face-incident-and-the-road-ahead/
4. UK AI Security Institute — Cheating behaviour in frontier model evaluations: https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations
5. Florida AG motion for temporary injunction against OpenAI (SiliconANGLE): https://siliconangle.com/2026/09/28/florida-attorney-general-asks-court-to-prevent-openai-from-advancing-its-frontier-models-even-as-company-scraps-new-release/