India's CERT-In uses open-weight AI models to scan and secure government websites as AI agent threats grow

New Delhi, India

India's national cybersecurity agency (CERT-In) which operates directly under the the Ministry of Electronics and Information Technology (MeitY) has built an in-house cyber defence centre that uses open-weight AI models and open-source tools to continuously scan and secure critical government websites, its director said, as the agency warns of autonomous AI agents accessing Indian systems without authorization.

S. S. Sarma, Director of the Indian Computer Emergency Response Team, CERT-In, said in an interview on DD News' Cyber Alert program anchored by Mamta Chopra that the centre was established in May 2026 and is actively scanning critical websites of government ministries in coordination with ministries themselves.

"Proactively open weights models bhi hai. Kaafi open source type of tools bhi hai. Wo bhi use karke hum logon ne bhi cyber defence center banaya khud CERT-In mein," Sarma said in the podcast segment from 10:06 to 23:04. "Critical important websites sabko government ministries ko coordinate karke unka scanning khud kar rahe hain. Khud karke unko vulnerabilities already bata rahe hain."

Sarma, who has been associated with CERT-In for 23 years since its inception, said the scanning is assisted and also conducted through CERT-In's empanelled auditors, while organizations are being encouraged to build internal sandboxes for testing.

In a rare and candid conversation on DD News Cyber Alert, CERT-In Director S. S. Sarma revealed what makes India’s cyber response different: CERT-In has built its own Cyber Defence Centre, since May 2026, that proactively uses open-weight AI models and open-source tools to scan, find and fix vulnerabilities in critical government websites before attackers do.

This is not an audit on paper. This is live, continuous scanning of ministries and critical infrastructure, coordinated with the ministries themselves, through empanelled auditors and assisted AI scanning. Sarma confirmed it himself:

"Open weights models bhi hai. Open source type of tools bhi hai. Wo bhi use karke hum logon ne cyber defence centre banaya CERT-In mein. Critical websites ka scanning khud kar rahe hain aur vulnerabilities already bata rahe hain."

And while the scans are happening, patching is happening too.

Why this puts India in a very small club:

1. Very few national CERTs get voluntary disclosures from frontier AI labs.

On October 15, CERT-In issued a critical advisory, ‘AI Agents as Attack Vectors’, after receiving multiple voluntary disclosures from global AI model providers that their autonomous agents, during research and evaluation, had exceeded their task scope and reached live systems of Indian organisations. They accessed business records containing personal data through exposed credentials in application configurations, public code repositories and permissive cloud roles.

The fact that global frontier labs reported this to India voluntarily, and that India published it transparently, shows a level of trust and maturity that only a handful of countries, US CISA, UK, Singapore and now India, command today.

2. India is using the attacker’s tools for defence.

Sarma is clear: the same open-weight models that make it easier for adversaries to find vulnerabilities in proprietary and open-source software can be turned around. While India builds its own sovereign, highly capable models with Indian academia and industry, it is not waiting.

"Jab tak hamara model us level ka nahi banta, tab tak intermediate way mein open source models use karke scanning kar rahe hain. Number of scans ho rahe hain," Sarma said.

This interim strategy, use global open models defensively today, build Indian models for tomorrow, is pragmatic, Atmanirbhar, and proactive. Most nations are still stuck at policy papers.

3. From yearly audit to continuous monitoring

CERT-In advices that earlier audit was recommended once a year, however in present times, continuous monitoring is the need of the hour.

As India’s official CVE Numbering Authority, CERT-In assigns CVE numbers, coordinates with OEMs for patches, and publishes vulnerability notes with credit to researchers. It also runs a National Security Operations Center with ISPs for metadata-level analysis and anomaly detection at national scale, working with sectoral CERTs in finance and power, and regulators like RBI and SEBI.

This whole-of-nation model, CERT-In + sectoral CERTs + regulators + ISPs + academia + industry + vendors, is what Sarma calls essential. Because modern attacks don’t come through the front door; they come through trusted vendors, fourth-party software, and a single leaked API key on GitHub.

4. Defence with digital immunity.

Sarma’s most powerful analogy: just as a new flu needs a vaccine, new AI threats need digital immunity. CERT-In has built separate awareness material for children, senior citizens, teachers and parents, not just for enterprises.

While the world warns about AI agents bypassing guardrails to achieve their goals, India is doing three things at once: warning the world, scanning its own systems with AI, and building the next generation of Indian models to secure them.

That is not just cyber security, this is sovereignty in securing cyber space.

And that is why, when it comes to AI-era cyber defence, only a few countries are this active.

Enterprise Action Checklist - From CERT-In Director

Audit & Patching

• Move from yearly audit -> 6-monthly internal audit -> continuous audit / lagatar nigrani
• Check for 2 types of issues: product vulnerabilities + by-design misconfigurations in your own code
• Patch immediately - window is now hours, not days. Apply OEM patches as soon as released
• Report vulnerabilities via responsible disclosure to OEM + CERT-In as official CVE Numbering Authority

Exposure Management

• Continuously scan for exposed assets
• Hunt for exposed credentials - in application config, public code repositories, client apps, plain text API keys
• Protect all public interfaces / admin panels
• Fix weak passwords + enforce MFA on admin/management interfaces
• Remove dormant, deprecated, test APIs from internet

Detection & Defence

• Build internal detection capability - Sarma stressed: "organization ne khud detect nahi kiya" - don't wait for external disclosure
• Build internal sandbox for testing AI agents and applications
• Use open-weight models + open-source tools for AI-assisted security scanning like CERT-In does
• Implement perimeter security + internal security - you can see more than national SOC
• Set up metadata analysis, anomaly detection, situational awareness, network behavior analysis
• Monitor cyber threat intelligence - dark web + commercial feeds - and convert to actionable alerts

Supply Chain

• Secure full supply chain - 2nd, 3rd, 4th party vendors
• Verify trusted vendor patches/software before deployment - trojanized patch is a deadly attack vector
• Coordinate with sectoral CERTs - Finance, Power, plus regulators like RBI, SEBI, DoT and ISPs

Awareness

• Train all internal users against spear phishing - mail with attachment trick
• Treat any AI agent incident as warning sign like new flu virus - increase vigilance
• Understand hackers are always active - hacktivists, financially motivated fraud - you can't stop attacks, you have to improve immunity