I work at the intersection of Applied AI and security engineering — red teaming and building controls for production AI systems at my workplace before they reach end users. Traditional security reviews don't cut it as a one size fits all for AI. My job is to find the things that change deployment decisions primarily through inference level simulations
What I'm working on:
- Engineered a firm-wide Gen AI Red Teaming toolkit enabling systematic automated and manual adversarial evaluation of LLM use cases across all lines of business
- Led the red team assessment of flagship LLMs, surfacing multiple critical and high-severity vulnerabilities — prompt injection, data leakage, system prompt extraction, and intent drift — before the models reached production
- Authored a blueprint for Advanced Manual Red Teaming: a structured methodology for uncovering behavioral vulnerabilities and intent misalignment that automated scanners fundamentally cannot detect
- Architected a Network Broker enabling secure, isolated connectivity between internal red teaming platform and external vendor evaluation infrastructure
- Containerized and deployed Red Teaming ECS Task modules for concurrent adversarial evaluation of multiple AI use cases without capacity constraints
- Reviewed emerging AI coding tools from a security perspective, producing key findings that directly shaped firm-wide decisions on AI coding enablement
- Building AI Controls engineering solutions to systematically enforce safety, scope, and behavioural boundaries across deployed LLM use cases
Recognition: Inventor Recognition (Q4 2025) for filed patents · Speaker at DEVUP 2026 (my workplace's invite-only technical conference) · SEP Engineer Committee Lead for 1,100+ early-career engineers at my workplace Bengaluru Tech Centre
Full-fledged engineering projects: red-teaming tools, safety benchmarks, and taxonomies built to address real gaps in how deployed LLMs are secured.
|
Interactive AI Security Hacking Playground A gamified Capture The Flag platform designed to teach hands-on adversarial thinking. Mapped directly to real-world risk frameworks (OWASP LLM Top 10 & MITRE ATLAS).
|
Prompt Injection Benchmark for Agentic Tool-Use The first benchmark targeting injection attacks in agentic, tool-using pipelines — tool-output injection and goal hijacking across multi-step workflows, where single-turn benchmarks like AdvBench and HarmBench don't reach.
|
|
Absolute Safety Robustness Evaluation Harness An advanced LLM safety benchmark that evaluates absolute, severity-weighted category failure rates instead of shifting, relative statistics (Z-scores).
|
Multi-Turn Jailbreak Taxonomy & Detection A taxonomy of trajectory-based LLM jailbreak patterns — gradual drift, persona anchoring, trust escalation — that single-turn automated scanners structurally can't catch, plus concrete detection methods.
|
|
AI Scope & Intent Enforcement Proxy Gateway An enterprise-grade safety proxy and automated red-teaming tool designed to keep deployed LLM applications safe, secure, and strictly aligned within their defined scopes.
🔒 Stealth / Private Repository |
The Definitive AI Security Practitioner's Guide A curated, practitioner-depth guide to AI red teaming, runtime security, and MLSecOps — 14 specialized handbooks, not a link dump.
|
prompt-injection-ctf (⭐ 6) — Interactive AI Security Playground — Prompt Injection CTF. Craft attack prompts to break constrained AI systems. Learn prompt injection, jailbreaking, intent drift & token smuggling. Built to teach adversarial thinking hands-on.
ai-security-resources (⭐ 4) — Curated directory of state-of-the-art Adversarial AI Security tools, vulnerability scanners, safety benchmarks, guardrails, and compliance standards.
AgentInjectionBench (⭐ 3) — The first benchmark targeting prompt injection attacks in agentic tool-use pipelines — tool output injection, goal hijacking, and multi-step attack chains. Dataset & Space on Hugging Face.
weighted-safety-refusal (⭐ 1) — Severity-weighted LLM safety evaluation suite. Measures absolute refusal robustness across prompt injection, jailbreaking, data exfiltration, toxicity, and malware generation using risk-adjusted weights.
synaptic-wetware — 🧠 Organoid Intelligence Biocomputer Simulator — HH + Izhikevich neuron models, MEA burst detection, DishBrain Pong, Baltimore Declaration ethics monitor. Built by Antigravity (Google DeepMind).
intent-drift-playbook — A taxonomy of trajectory-based LLM jailbreak patterns (gradual drift, persona anchoring, trust escalation) that single-turn automated scanners structurally can't catch — plus concrete detection methods.
ai-security-tracker — Real-time security monitoring across 50+ AI/ML repositories — tracks vulnerabilities, CVEs, and security initiatives ecosystem-wide.
llm-ops-workshop — End-to-end MLOps workflow demonstrating model lifecycle, monitoring, and deployment practices.
Also building something in AI security — stealth mode 🔒
📋 Full list of external contributions — repos outside my account where I've contributed (2020–present)
LLMs & AI Platforms
Red Team & Security Tools
Frameworks & Standards
Tech Stack
