Minjae Gwon

Won DARPA’s AI Cyber Challenge Final as part of Team Atlanta (Georgia Tech, Samsung Research, KAIST, POSTECH). Built components of an autonomous system that detects and patches vulnerabilities in open-source code at competition scale.

Designed and trained a multi-turn retrieval agent for security patching. Optimized GRPO-based reinforcement learning rewards so the model selects minimal, high-value code context for downstream patch generation.

Led development of the end-to-end patching framework. Maintained a typed monorepo with dependency-injection patterns, caching, and reproducible runs; integrated evaluation harnesses and dataset tooling; and ran codebase-wide reviews to enforce clean abstractions and fault isolation.

Won the DARPA AI Cyber Challenge (AIxCC) Semifinals at DEF CON 32 in Las Vegas. Contributed to automated patch generation with large language models (LLMs), enabling the system to detect and patch software vulnerabilities autonomously across multiple programming languages.

Built ChainLight DART, an autonomous Web3 threat analysis platform with a Rust-based core. Led engine development, integrated continuous data processing components, and introduced engineering practices including monorepo workflows, stricter code quality standards, and stronger type safety.

Contributed to Relic Protocol, a secure on-chain historical data source built with zero-knowledge proofs. Developed front-end applications, reviewed smart contracts and SDKs, and built productivity tooling including wiki-to-Markdown conversion scripts.

Evaluated a model that mitigates adversarial attacks by extending prior work such as MagNet and SHIELD.

Researched automated vulnerability detection using differentiable line-probability metrics and prefix tuning on CodeLlama 7B to improve line-level vulnerability identification.

Analyzed client codebases and identified critical business-logic vulnerabilities.

Developed and reviewed training content for Dreamhack Enterprise, including exercises on HTTP and file-handling security. Also authored a Dreamhack CTF challenge focused on UUID1 and SSTI exploitation.

Built a blockchain-based payment system that won first prize. Led front-end development and beta testing. The project was later selected for the ICON Foundation Grant Program.

Competed in CTF events as a core web-hacking member and delivered internal seminars on practical offensive security techniques.

Mentored high school students in web security from fundamentals to advanced exploitation. Built hands-on attack/defense training environments and designed multiple CTF challenges.

Led media production for the camp, creating branded assets and operating live broadcasting systems for seminars.