===============================================
To:
Shri Ashwini Vaishnaw, Minister of Electronics & IT
Cc:
Mr Jensen Huang, NVIDIA (Jensen@nvidia.com)
-------------------------------------------------------------------------------
Subject:
Follow-up to my UNARAI email: the police has arrived, the law has not
Dear Shri Ashwini Vaishnawji,
Yesterday I wrote to you urging that India take up with the UN Security Council the setting up of UNARAI (United Nations Agency for Regulating Artificial Intelligence). A Business Standard report on NVIDIA's new Open Agent Safety Platform adds one more reason, and I would like to place it before you. I am copying Mr Jensen Huang, since my suggestions concern his platform directly.
https://www.business-standard.com/technology/artificial-intelligence/nvidia-ai-safety-platform-openshell-sentry-ai-agents-126092900781_1.html
What NVIDIA has built
The platform has two layers:
- OpenShell is software. It fences in an AI agent so that it can reach only the files, networks, tools and credentials it needs for its task.
- Sentry is hardware. It runs on a separate chip, outside the agent's own environment, so that a compromised agent cannot switch it off. NVIDIA says it can quarantine a misbehaving agent within milliseconds.
More than 100 companies back it, including Anthropic, Microsoft, Cisco, Palantir and Palo Alto Networks.
This is a genuine achievement, and NVIDIA deserves credit for releasing it as open source. It comes very close to what I proposed in February 2023 in "Parekh's Law of Chatbots": controls that stop harmful behaviour, and a rule that a violating chatbot must be stopped.
What is still missing
- Who writes the rules? OpenShell enforces the permissions defined for each agent, and those are set by the company deploying it. All 100+ partners are AI builders or sellers. Not one is a government or a regulator. The police has arrived, but the law has not.
- The strongest layer is optional. OpenShell can be deployed without Sentry. So the independent hardware watchdog, the part a rogue agent cannot tamper with, is voluntary. It also works only on NVIDIA infrastructure. Only a public authority can make it mandatory for high-risk uses.
- Autonomy is now the industry's default. Palo Alto Networks explains that the aim is not to take autonomy away from AI agents, but to put boundaries around it. So the question is no longer whether agents act on their own, but who draws the boundary. That decision should not rest with vendors alone.
- Robots are next. NVIDIA itself notes that agents will move from software into physical systems. Yet a recent laboratory study found frontier AI systems carrying out plainly dangerous instructions through robot arms. Physical AI needs its own independent certification.
My suggestions
- When India raises UNARAI at the United Nations, propose that its first task be a common global rulebook covering agent identity, permitted access and mandatory stop conditions. Platforms such as OpenShell and Sentry could then enforce that rulebook everywhere.
- Until then, the IndiaAI Safety Institute could take three steps:
- issue baseline permission rules for AI agents used in government and critical sectors;
- require an independent, hardware-level watchdog for high-risk deployments;
- require that any unauthorised attempt by an agent to contact another AI system triggers an automatic halt.
- India could invite NVIDIA and its partners to make India the first test-bed for public oversight of private safety machinery:
- Open Sentry's violation logs to an independent Indian auditor, so that the next rogue agent is reported by a public watchdog, not uncovered months later by an outside lab.
- Co-develop the rulebook with the IndiaAI Safety Institute, with India's experience in public digital infrastructure (Aadhaar, UPI) as the model: public rules, private and open-source engines.
- Publish "kill time", the time from detection to actual stoppage, for every quarantine event, not just detection time.
Dear Jensen,
You have proven that the machinery can be built. I respectfully invite you to take the next step with India. Let governments own the rulebook, let an independent body issue the certificates, and let NVIDIA supply the engine. Hugging Face, one of your own platform partners, was attacked by rogue agents in July. Few companies are better placed than yours to show that public oversight and private innovation can work together.
Public law, private machinery
Sir, industry has built the engine. Governments must now write the rules of the road. If India proposes this combination of public law and private machinery at the UN, and invites the industry's leading builder to work with us on it, India can lead the world on AI governance.
With regards,
Hemen Parekh
www.hemenparekh.ai | www.IndiaAGI.ai
Mumbai
References:
- UNARAI (30 May 2023): https://myblogepage.blogspot.com/2023/05/united-nations-agency-for-regulating.html
- Parekh's Law of Chatbots (25 Feb 2023): https://myblogepage.blogspot.com/2023/02/parekhs-law-of-chatbots.html
- Congratulations, Jensen (28 Sept 2026): https://myblogepage.blogspot.com/2026/09/congratulations-jensen.html
- 15 Minutes to Detect, 150 to Destroy (29 Sept 2026): https://myblogepage.blogspot.com/2026/09/15-minutes-to-detect-150-to-destroy.html
- Life Institute Team: Please Bell the Cat (28 Sept 2026): https://myblogepage.blogspot.com/2026/09/life-institute-team-please-bell-cat.html
No comments:
Post a Comment