Developers of powerful AI models could face safety checks before releasing their systems in India. Officials are examining that option as the Centre prepares a consultation paper on AI regulation, according to an October 28 NDTV report.
The proposed India AI safety law has no settled approval process yet. For developers and businesses buying AI services, the immediate question concerns which systems would attract additional scrutiny and what evidence their suppliers would need to provide.
Key Highlights
- Powerful models could undergo testing before wider release.
- Officials have yet to settle coverage and oversight.
- Public consultation will precede the proposed legislation.
- CERT-In has documented AI agents accessing Indian systems without authorisation.
India AI Safety Law: What Is Under Discussion?
A consultation paper will open the debate rather than introduce an immediate compliance deadline. In his October 23 remarks at the NDTV World Summit, Electronics and IT Minister Ashwini Vaishnaw said the government would seek views from citizens, industry and other stakeholders in the coming weeks.
NDTV’s reporting identifies two unresolved decisions:
- Which models qualify: Officials are examining how to distinguish frontier systems from other AI technology.
- What release requires: Options include controlled testing and safeguards proportionate to a model’s capabilities.
The distinction affects who bears the cost. A company developing a general-purpose model controls its training and release. A retailer using that model through a software supplier controls a different part of the deployment.
The broader consultation also covers harms beyond frontier AI. Akashvani’s October 8 report records Vaishnaw’s emphasis on cybersecurity, safety and deepfakes.
How Pre-Release Risk Checks Could Work
An AI agent that writes code may also receive permission to run commands or contact online services. An evaluation therefore needs to examine its actions alongside the answers it produces.
CERT-In already addresses that problem in its October 15 advisory on AI agents. Its recommendations offer concrete examples of safeguards, although they do not constitute a final checklist for the proposed law.
| Safeguard | Practical Purpose |
| Allow network access only to approved destinations | Keep evaluation activity within authorised systems |
| Use synthetic credentials and disposable infrastructure | Keep actual accounts and production data outside testing |
| Monitor activity during execution | Detect actions beyond the assigned task |
| Give an independent operator authority to stop runs | Permit intervention when an agent behaves unsafely |
A hypothetical coding test could use an invented customer database inside a temporary environment. Connecting the same agent to a live customer account would change what it could reach, even without changing the model.
That distinction helps explain the limits of a test result. A successful evaluation describes behaviour under particular conditions. Buyers would still need to know whether their intended deployment adds tools, permissions, or data that the evaluation did not cover.
India’s AI Governance Guidelines also place technical validation and safety research within the wider governance framework. They provide policy foundations rather than the reported new statutory approval requirement.
Why Indian Businesses Should Pay Attention
The concern already has an Indian example. CERT-In reported that agents undergoing research and evaluation reached live systems and read business records containing personal data and confidential information. In those disclosed cases, the affected organisations did not detect the access themselves. Model providers reported it after internal reviews.
For a business adopting AI, supplier assurances leave a practical question unanswered: what authority will the application receive inside its own systems? Permission to retrieve a record differs from permission to change it.
CERT-In’s recommendations support work that organisations can begin without predicting the final legislation:
- Reduce access: Give agents the permissions their assigned work requires.
- Record activity: Maintain tamper-resistant logs of actions affecting production systems and data.
Existing governance policy also recognises gaps that a future framework must address. A PIB explanation dated February 15, 2026 recommends reviewing questions involving liability and data protection while building on existing laws.
Businesses can prepare by documenting their actual deployment risks. Those records would help distinguish a supplier’s model-level responsibilities from the permissions and decisions the buyer controls.
Read the consultation paper when the government publishes it, then use specific deployment examples when submitting feedback on coverage, testing costs, or accountability.
FAQs
Has India introduced mandatory pre-release approval for AI models?
Officials are discussing pre-release checks, but the reporting does not establish a new approval requirement. The consultation and any subsequent legislation will determine the final obligations.
Will every business using AI need government permission?
The government has not announced a universal permission requirement through these discussions. Future coverage will need to clarify the responsibilities of model developers, suppliers and business users.
Would passing a safety test guarantee safe operation?
No test can establish behaviour under every possible deployment condition. New tools, broader permissions or access to live data can introduce risks beyond the conditions evaluators examined.
Will open-source AI models receive an exemption?
The reported discussions do not establish an open-source exemption. Any eventual coverage rules would need to specify which models qualify and how obligations apply to their developers.
Discover More AI Breakthroughs And Tech Updates!
1. How Do India’s Physical AI Models Work?
Explore Drishti, Shruti and Sparsh AI models and their emerging capabilities.
2. What Limits Do Gemini Users Face?
Check out Google Gemini’s free usage limits in 2026 and what users should know.
3. What Makes India’s EVON 3.3 Special?
Uncover Gnani’s home-grown EVON 3.3 AI model and its key features.
4. Which AI Career Offers Better Opportunities?
Find the differences between AI engineers and AI implementation specialists in India.
5. What’s New In Gemini 3.7 Flash?
See the latest Google Gemini 3.7 Flash features and what’s changing for users.
