AI security engineer. I took a product from zero to production and 40+ paying clients in a month, and I build the controls that stop agents from doing damage once they have real tool access.
What I'm proudest of is an open-source MCP security gateway (Inline Proxy). It sits between an agent and its tools and enforces policy on every call: allow, block, redact, require approval. The case it exists for: a poisoned GitHub issue tells the agent to publish the repo's .env as a public gist. The gateway never has to detect the injection. It only tracks that untrusted content was read, so the exfiltration write gets blocked. Default-deny packs cover 194 tools across GitHub, Jira and Slack, ~0.1ms p99 on the enforcement path, fail-closed. [Inline Proxy repo link]
My master's thesis (UW 2026, ProQuest 32738510) became AegisLLM, an LLM red-team suite mapped to the OWASP LLM Top 10 that uses a genetic algorithm to evolve new attacks. github.com/Harshj143/AegisLLM
MS in Cybersecurity Engineering (UW), top 1% on TryHackMe, CTF win at UWB GreyHats. Looking for security engineering or AI security roles at companies putting agents in production. Happy to walk through the design of any of this.
Remote: Yes (remote or hybrid)
Willing to relocate: Yes (SF Bay Area, Seattle, NYC, Austin and other US hubs)
Technologies: Python, security engineering, agentic AI and MCP security, LLM red-teaming, guardrails, threat modeling, DevSecOps, AWS, Docker, Next.js, Node.js
Résumé/CV: https://www.harshjannawar.com/resume | github.com/Harshj143
Email: [email protected]
AI security engineer. I took a product from zero to production and 40+ paying clients in a month, and I build the controls that stop agents from doing damage once they have real tool access.
What I'm proudest of is an open-source MCP security gateway (Inline Proxy). It sits between an agent and its tools and enforces policy on every call: allow, block, redact, require approval. The case it exists for: a poisoned GitHub issue tells the agent to publish the repo's .env as a public gist. The gateway never has to detect the injection. It only tracks that untrusted content was read, so the exfiltration write gets blocked. Default-deny packs cover 194 tools across GitHub, Jira and Slack, ~0.1ms p99 on the enforcement path, fail-closed. [Inline Proxy repo link]
My master's thesis (UW 2026, ProQuest 32738510) became AegisLLM, an LLM red-team suite mapped to the OWASP LLM Top 10 that uses a genetic algorithm to evolve new attacks. github.com/Harshj143/AegisLLM
MS in Cybersecurity Engineering (UW), top 1% on TryHackMe, CTF win at UWB GreyHats. Looking for security engineering or AI security roles at companies putting agents in production. Happy to walk through the design of any of this.