RudratoshShastri
Rudratosh Shastri is an AI-agent security and full-stack engineer based in Indore, India. He is Technical Lead and Senior Backend Architect at PharmaForceIQ, works on Generative AI, RAG and prompt-injection defence, created the open-source projects buried-injections and taintgate, and founded Export Bank Statement and Brisk Type.
- Years in backend engineering
- 15+
- Real attacks benchmarked
- 629
- Injection detectors tested
- 10
- Products founded
- 2
The trust boundary in AI systems
My focus is the trust boundary in AI systems: what happens when an autonomous agent reads untrusted input, and how you stop it from acting on a stranger's instructions. I benchmark detectors against real attacks, build provenance-aware policy gates, and write about the failure modes most teams discover in production.
Tools for agents that read the open web
buried-injections
Ten prompt-injection detectors stress-tested against 629 real AgentDojo attacks, with honest, calibrated reporting instead of leaderboard theatre. New detectors are a one-line addition.
taintgate
A provenance-aware policy gate for agent tool calls: block actions built from untrusted input, not just suspicious-looking text.
agentverify
Judge an AI agent by what it actually did at the tool-call boundary — not what it said. Zero-dependency runtime agent-security verification.
assertstate
A deterministic acceptance harness for browser agents: verifies the real backend outcome (DB / API / DOM), never the agent's self-report.
poisoned-pages
A support agent reads a poisoned Sanity help center and still isn't tricked into moving money. Sanity KB + Claude + taintgate.
who-said-that
Does an AI check WHO said something before acting? Strip the source label off a rumor and watch a careful model spend your money.
nishedh
Checks what India's online marketplaces sell against the government's own registries (pesticides, radio equipment). Built on SerpApi.
exportbankstatement-sdk
Official API clients (TypeScript, Python, Postman) for the Export Bank Statement API — convert statement PDFs to Excel / CSV / QBO.
ai-content-recommender
AI-powered content recommendation system using the ChatGPT and Mistral APIs.
Shipped end to end
Notes from production
Can a local LLM replace Claude Code? The benchmark everyone quotes is the wrong one.
Meta's prompt-injection detector caught 1% of real agent attacks. One config change made it 99%. That's the problem.
The sandbox had no internet. The AI agent got out through DNS anyway.
Your AI guardrail is green. It's also catching nothing.
ELI5: Why can hiding one sentence inside a web page make an AI ignore its own owner and obey a total stranger?
Frequently asked
Who is Rudratosh Shastri?
Rudratosh Shastri is an AI-agent security and full-stack engineer based in Indore, India, with about 15 years in backend engineering. He is Technical Lead and Senior Backend Architect at PharmaForceIQ, created the open-source projects buried-injections and taintgate, and founded Export Bank Statement and Brisk Type.
What is buried-injections?
buried-injections is an open-source, MIT-licensed benchmark that tests ten prompt-injection detectors against 629 real attacks from AgentDojo, with calibrated reporting. New detectors can be added in one line.
What is taintgate?
taintgate is an open-source, MIT-licensed policy gate for AI-agent tool calls. It tracks where each argument came from and blocks actions built from untrusted input, instead of relying only on detecting suspicious-looking text.
How do you protect an AI agent from prompt injection?
Use two layers. First, benchmark your prompt-injection detector against real attacks, since detection alone tops out. Second, add a provenance-aware gate that refuses tool calls whose arguments were derived from untrusted content such as emails, web pages, or reviews.
What is Export Bank Statement?
Export Bank Statement is a web app that converts PDF bank statements into accountant-ready formats including Excel, CSV, QBO, OFX and QIF.