Most LLM-based products ship with no input security layer. We know because we found and responsibly disclosed vulnerabilities in Le Chat (Mistral) and Sarvam AI's Indus model — system prompt extraction, guardrail bypass, phishing content generation. Rakshak sits between your users and your model. Blocks on the way in, redacts on the way out. Built for Indian AI teams but works with any LLM API. Self-serve API key, integrates in under an hour.

Mintlify — Knowledge infrastructure for AI agents

Knowledge infrastructure for AI agents

Hey PH 👋 I'm Shubham, founder of KalpitLabs. Rakshak started from a simple observation — we were red teaming Indian AI products and finding the same vulnerabilities everywhere. System prompt extraction, prompt injection, guardrail bypass. We disclosed them in Le Chat (Mistral) and Sarvam AI's Indus model. Both patched after our reports. The problem wasn't that these teams were careless. It's that there's no standard input security layer for LLM deployments — especially in India where English-only guardrails miss Hindi and Urdu attack vectors entirely. So we built Rakshak. Three-stage pipeline — pattern detection, semantic similarity, LLM classifier. Sits in front of any LLM API. One integration, blocks what goes in, redacts what leaks out. Would love feedback from anyone building LLM-based products — especially around edge cases you've hit that existing guardrails miss. Happy to answer anything. 🙏

Mintlify — Knowledge infrastructure for AI agents

Knowledge infrastructure for AI agents

Rakshak by KalpitLabs

LLM guardrail API. Stops before it hits your model.

LLM guardrail API. Stops before it hits your model.