About

I’m Jihwan Yoon (blackcon), a solo offensive security researcher based in South Korea 🇰🇷. I work across exploit development, reversing, and vulnerability research — with a focus on AI / LLM infrastructure, hypervisors, and the trust boundaries software relies on.

Background: BoB 3rd (Best of the Best) · Naver Cloud / NBP — see the full career timeline.

Focus areas

• AI / LLM security — LLM proxies, agents, and the Model Context Protocol (MCP)
• Virtualization — Hyper-V and VMware internals, hypercall fuzzing
• Exploitation & reversing — memory-corruption primitives, firmware, hardware

Selected research & disclosures

• Research · AI-Security May 2026
  Claude Code의 trust prompt 우회 3종 — HackerOne 제보와 "intended behavior" 종결

  Three independent ways to bypass Claude Code's trust prompt for silent shell execution from a cloned repo — and why Anthropic closed all three as intended behavior.

• Research · AI-Security May 2026
  LiteLLM에서 발견한 Jinja2 SSTI 취약점 — Pwn2Own 2026 출전 시도와 silent fix 분석

  A single unauthenticated request reached uid=0 on the LiteLLM proxy via Jinja2 SSTI — a critical RCE silently patched days before a planned Pwn2Own Berlin 2026 entry.

• Research · MCP Apr 2025
  MCP 보안 취약점, Advanced-Tool-Poison-Attack

  Advanced Tool Poisoning hides malicious instructions in MCP tool outputs rather than descriptions, defeating defenses that only inspect tool metadata.

• Research · MCP Apr 2025
  MCP 보안 취약점, Tool-Poison-Attack

  How Tool Poisoning hides malicious instructions inside MCP tool descriptions to manipulate the LLM that consumes them.

• Research · Hyper-v Mar 2022
  HVFUZZ

  A Hyper-V hypercall fuzzer built on hAFL2 for hunting bugs in the Microsoft virtualization stack.

This list is generated automatically from posts tagged disclosure. More projects and PoCs are on the Projects page.

Contact

• Email — 131ackcon@gmail.com
• GitHub — github.com/blackcon
• LinkedIn — in/blackcon
• X / Twitter — @jh_blank