Harsh Jannawar, Security Engineer

§ 01

About

Subject file

I ship security tooling, not just findings.

Security Engineer who builds secure, resilient systems and ships security tooling rather than just finding problems. Experienced across application security, detection engineering, cloud, and AI/LLM security, aligned with OWASP, MITRE ATT&CK, and NIST.

Most of what I build answers one question: how do you let an AI system be genuinely useful without letting it be dangerous? That has become a transparent MCP proxy that enforces per-tool policy on every agent call, a red-team suite mapped to the OWASP LLM Top 10, and a SOC assistant that writes incident reports from playbooks and raw logs.

At SecureAIs I ran adversarial testing against production AI platforms and found 20+ critical issues including authentication bypass, prompt injection, and token leakage before any of it shipped, then built the PII detection and redaction pipeline that pushed sensitive-data protection accuracy from 65% to 95%.

At SecureThings I spent a year on the build side: an event-driven scanning platform in Dockerized async Python that cut asset analysis from nine hours to under thirty minutes, backend risk-scoring logic to prioritize what actually mattered, and AWS/Azure hardening across 25+ security gaps.

Off the clock: CTFs with UWB GreyHats, cloud hardening rabbit holes, and writing up what breaks.

Education

University of Washington

M.S. Cybersecurity Engineering

Sep 2024 – Jun 2026

GPA 3.93 / 4.00

Symbiosis Skills and Professional University

B.Tech, CSIT (Cybersecurity)

Aug 2020 – May 2024

GPA 3.65 / 4.00

Capabilities

Security

Security tool development
Threat modeling
Application security reviews
Penetration testing
Secure code review

Languages

Python
JavaScript
Node.js
Bash

Frameworks

ISO 27001 / 42001
NIST
SOC 2
OWASP
MITRE ATT&CK

Tooling

Burp Suite
Metasploit
Nmap
Nuclei
Splunk
Docker

Cloud

AWS
Azure
Lambda
S3
CloudTrail
GuardDuty

Data

MongoDB
Postgres
SQL

§ 02

Experience

2022–2025

Jun 2025 – Aug 2025

SecureAIs

California, USA

AI Cybersecurity Engineering Intern

01Developed a PII detection and redaction pipeline using Presidio, regex-based validation, and packet analysis to protect sensitive data across 10k test samples, improving sensitive-data protection accuracy from 65% to 95%.
02Conducted adversarial testing and secure code reviews across AI platforms to uncover 20+ critical vulnerabilities including authentication bypass, prompt injection, and token leakage prior to production release.
03Integrated SAST, DAST, dependency scanning, and secret detection into CI/CD pipelines to give continuous security feedback during development, reducing security review cycles by 60%.
04Built and tested end-to-end authentication flows across engineering services, covering secure login, session handling, token validation, role-based access control, and protected API routes, to reduce unauthorized access risk across application workflows.

PresidioAdversarial TestingSAST / DASTCI/CDAuthN / AuthZ

Jun 2023 – May 2024

SecureThings

Pune, India

Security Analyst Intern

01Built an event-driven scanning platform using Dockerized asynchronous Python, Nmap, and Nuclei, deployed via Amazon ECR and ECS, persisted to S3 and MongoDB, and triggered by AWS Lambda on new user requests, cutting asset analysis time from 9+ hours to under 30 minutes.
02Performed application and infrastructure security assessments using Nessus, Nmap, and Burp Suite to triage 15+ high-severity vulnerabilities with validated exploits.
03Designed backend risk-scoring logic to prioritize findings by exploitability, exposed services, affected technologies, severity, and business context for faster remediation decisions.
04Strengthened AWS and Azure environments by improving IAM permissions, container security controls, secrets handling, and exposed service configurations, remediating 25+ cloud and application security gaps.

Python asyncioDockerAWS ECS / LambdaNucleiBurp SuiteMongoDB

Jun 2022 – Jul 2022

Pune Metro Rail

Pune, India

Security Intern

01Reviewed IAM roles and security group configurations across 20+ cloud resources to identify over-permissive access and exposed services, recommending least-privilege changes that reduced the attack surface across 15+ findings.

AWSIAMCloud Review

Full history in the résumé, linked below.

§ 03

Projects

8 projects

AI is part of the attack surface now. These are the tools I built for it.

2026–Present

Security Gateway for AI Agents

Inline Proxy

A transparent MCP proxy that enforces policy on every tool call an agent makes.

Role-aware policy on every tool: allow, block, redact, rewrite, quarantine, or route to human approval. The decision happens before an action runs rather than after. Fail-closed by default, with a pluggable asyncio pipeline and layered YAML policy packs.

PythonasyncioMCPYAML PolicySIEM

▪Per-tool policy decisions on every agent call, fail-closed by default
▪Defense-in-depth against prompt-injection exfiltration: PII/secret redaction with checksum-validated detectors and a tokenization vault
▪Taint tracking, per-session risk scoring with auto-suspend, and an LLM behavioral anomaly monitor
▪200+ tests in CI, with connector packs for GitHub/Jira/Slack, OIDC identity via Okta/Auth0, policy-as-code CI/CD, and audit streaming to Splunk/S3

2026

LLM Red-Team Suite

AegisLLM

Automated security testing for LLM applications, mapped to the OWASP LLM Top 10.

Runs thousands of injection and jailbreak prompts against a target, scores each one with LLM-as-a-Judge, and diffs the result against baseline models so a finding is a regression you can act on rather than a screenshot.

PythonFlaskOWASP LLM Top 10

▪Thousands of injection and jailbreak prompts, scored by LLM-as-a-Judge
▪Differential testing against baseline models to separate real findings from noise
▪ThreatForge, a genetic-algorithm prompt-evolution engine that breeds new attacks
▪Recon module fingerprints the target model, detects RAG, and enumerates connected tools

2026

Incident Investigation

SOC RAGBot

A SOC assistant that writes citation-aware incident reports from playbooks and raw logs.

Retrieval-augmented investigation for security analysts. It combines institutional knowledge with live log data and refuses to answer beyond what its sources support, so every claim binds back to where it came from.

PythonRAGLLM Ops

▪Citation-aware retrieval with enriched metadata and source binding to prevent hallucination
▪Semantic caching to cut latency on repeated investigations
▪Deterministic guard layer for prompt injection and jailbreak detection
▪Enterprise audit and replay layer capturing full interaction logs

04Detection Engineering2026

Cloud Detection Engine

Three-layer cloud threat detection modeled on the 2019 Capital One breach.

Ingests CloudTrail, WAF, and VPC flow logs, runs ten detection modules over them, scores entity risk, then explains findings for three different audiences and maps each one to a compliance control.

PythonCloudTrailMITRE ATT&CKPCI-DSS

05Email Security2026