HARSH VYAS

I test AI chatbots, LLM-powered apps, and AI agents for security

weaknesses. If you've deployed a chatbot, a RAG app, or any

agentic AI, it likely has gaps that traditional web pentesting

won't catch.

What I test for (OWASP LLM Top 10):

- Prompt injection — direct and indirect (hidden instructions in documents, web pages, or tool outputs)

- Jailbreaks and guardrail bypasses

- System prompt and sensitive data leakage

- Insecure output handling — XSS, SSRF, or RCE via model output

- Excessive agency — agents performing unauthorized actions or tool calls

- RAG and vector database weaknesses (data poisoning, cross-user leakage)

- Denial-of-wallet attacks that spike your API bill

Methodology:

Manual red-teaming combined with tools like Garak, PyRIT, and

Promptfoo, guided by the OWASP LLM Top 10.

What you get:

A clear PDF report with each finding, working proof-of-concept

prompts, severity rating, and recommended mitigations guardrails,

input/output filtering, system prompt hardening, and tool

permission scoping. One free re-check after you apply fixes.

What I need from you:

- Written permission to test

- Access to the chatbot, API, or test account

- Brief on intended behavior — what it should do, what it should refuse, what tools or data it touches

Honest note:

This field evolves fast. I follow current research and frameworks,

but no AI system can be made fully "secure" only harder to break.

I'll show you the gaps and how to close them.

Turnaround: 3–7 days depending on complexity.