The problem
AI systems fail differently from ordinary software. A traditional security review will not find a prompt injection, a jailbreak path, or a model that leaks its context under pressure — yet these are precisely the failures that make headlines. Teams shipping LLM-powered products today largely test them the way they test CRUD apps, and attackers know it.
The approach
Kavach is Hindi for armor. KavachRT integrates three evaluation surfaces that are usually separate tools — or missing entirely — into one workflow:
- Adversarial testing. Systematic attack campaigns against the deployed AI system: prompt injection, jailbreak attempts, goal hijacking, and data-exfiltration probes — run repeatably, not as a one-off pentest.
- Code-level risk detection. Analysis of the application code around the model: unsafe prompt construction, unvalidated tool calls, over-broad permissions granted to agents.
- LLM vulnerability scanning. Structured scanning of model behavior against known vulnerability classes, producing findings a security team can triage like any other report.
The output is not a score for a slide deck — it is a concrete finding list: what broke, how it was broken, and what to change. The same philosophy as Mizaan applies here: an evaluation you cannot act on is theater.
Why this lab
This product sits at the intersection of the lab's two deepest roots: a security career spanning DevSecOps at the IAEA and peer-reviewed malware research (ArkThor, ICISSP 2024), and current hands-on work in AI red teaming. KavachRT is where that experience becomes a product.
Status
In active development. Follow progress at kavachrt.com.