Quick Take
- Government asks Sarvam and BharatGen to build Mythos-like cyber capabilities for critical infrastructure defence.
- CERT-In sandbox models already deliver 60 to 70 percent of Mythos capability, says MeitY.
- India still lacks access to Anthropic Mythos, with US export talks continuing right now.
In This Article
The Indian government has asked homegrown AI firms Sarvam and BharatGen to tweak their foundation models and build Sarvam BharatGen Mythos-Like AI Models that can defend critical infrastructure against cyberattacks, officials said on July 13, 2026.
The plan is to host these indigenous models on the government’s isolated compute infrastructure. They would perform the same code-probing tasks that Anthropic’s Mythos performs. India still does not have access to Mythos, and talks with the United States government continue. Both firms are backed by the IndiaAI Mission, MeitY’s Rs 10,371.92 Cr national AI programme.
StartupFeed Insight
The real signal here is not the model, it is the compute. Sarvam raised $234 Mn in June 2026 explicitly to train a frontier model for agentic, coding, and cybersecurity work. The government asking for cyber capability three weeks later is not a coincidence, it is a procurement pipeline forming in public. StartupFeed expects the first CERT-In sandbox results from a Sarvam or BharatGen cyber variant to surface by Q3 FY27, and expects at least one defence or power-sector pilot contract to follow within twelve months. Founders in Indian security tooling should watch this closely, sovereign AI is turning into a buyer, not just a grant. By Avinash.
Sarvam BharatGen Mythos-Like AI Models: The Numbers
Sarvam BharatGen Mythos-Like AI Models are indigenous foundation models being modified to detect and fix software vulnerabilities in India’s critical public assets. Both firms already sit inside MeitY’s foundation model track. Neither has shipped a dedicated cyber variant yet.
| Metric | Detail | Notes |
|---|---|---|
| Firms asked | Sarvam, BharatGen | Both IndiaAI Mission foundation model partners |
| Current capability gap | 60 to 70 percent of Mythos | CERT-In sandbox estimate, MeitY secretary S Krishnan |
| Hosting | Government isolated compute | On-premises preferred for sensitive code |
| BharatGen IndiaAI support | Rs 988.6 Cr | Largest allocation under the foundation model pillar |
| Sarvam Series B | $234 Mn first close | June 15, 2026, at $1.5 Bn valuation |
| Deployment deadline | Not disclosed | Officials say models should deploy soon |
The most striking number is the 60 to 70 percent figure. CERT-In built that capability from a mix of open-source models, without any frontier model licence. It is a working stopgap, not a solution.
About Sarvam and BharatGen
Sarvam, founded in 2023 by Vivek Raghavan and Pratyush Kumar in Bengaluru, builds full-stack sovereign AI, from Sarvam-30B and Sarvam-105B models to voice and vision systems. It raised $234 Mn in June 2026 at a $1.5 Bn valuation, backed by HCLTech, Bessemer, Khosla Ventures, and Peak XV. BharatGen is an IIT Bombay-led consortium of nine institutions, running the Param model family as a government-owned sovereign AI stack.
What does this mean for India cyber defence?
India’s cyber defence now runs on a substitute stack while the frontier model stays out of reach. CERT-In has built a sandbox, described internally as a war room, where open-source models probe code, flag security gaps, and push fixes. MeitY secretary S Krishnan confirmed the arrangement publicly on July 13, 2026.
Getting access to Mythos and similar advanced frontier models is very high on the priority of the government, and this is something that we have taken up with our counterparts in the US and with the respective companies, S Krishnan, Secretary, MeitY.
Krishnan added that highly sensitive code needs on-premises deployment inside India, not overseas cloud. That single condition rules out most foreign frontier models by default. It also explains why the government wants Sarvam BharatGen Mythos-Like AI Models running on its own isolated compute infrastructure, where the code never leaves the country. The government has already identified the critical public assets these models will scan.
How do Sarvam and BharatGen compare?
The two firms bring different assets to the same problem. Sarvam brings private capital, shipped frontier models, and a stated cybersecurity roadmap. BharatGen brings the largest public allocation and academic depth across nine institutions.
| Dimension | Sarvam | BharatGen |
|---|---|---|
| Structure | Private company, unicorn | Section 8 foundation, IIT Bombay-led |
| Flagship models | Sarvam-105B, Sarvam-30B | Param2 family |
| Cyber mandate | Named in Series B use of funds | New, added via this government ask |
Sarvam holds the edge here. Its Series B announcement already named cybersecurity as a target use case for its next frontier model, three weeks before this government request became public.
Why do experts call this a hard task?
Replicating Mythos-class capability is a compute problem before it is a research problem. Mythos found decades-old flaws in widely used software, including a 27-year-old denial-of-service bug in OpenBSD. That class of autonomous exploit discovery needs cluster-months of frontier compute, which India is still assembling.
The clock matters too. The average time to exploit an unknown vulnerability fell to 44 days in FY26 from 745 days in FY22, according to a DSCI-BCG report. CERT-In’s own Blueprint for Defending against AI-Assisted Exploitation, issued May 25, 2026, demands 12-hour containment on internet-facing systems. Indigenous models must close that gap, not just narrow it.
What’s Next
Officials say the tweaked models should deploy soon, without giving a deadline. Watch for a CERT-In sandbox update or an MeitY disclosure on cyber variant benchmarks in the coming quarters. The negotiation with Washington over Mythos access runs in parallel, and either track could move first. Can India build a Mythos-class defender before attackers get one?
Frequently Asked Questions
Written by Avinash. Have a tip? Write to us at editorial@startupfeed.in.
