AI Security Researcher
Reproducible tools for authorized testing — turning attack paths into defensive evidence.
flowchart LR
A[Prompt injection] --> B[Agent hijacking] --> C[MCP exploitation] --> D[Tool poisoning]
Four links in one chain. I work on LLM and agent red teaming, MCP and tool security, secret and token exposure, and the defensive automation that turns findings into repeatable evidence: CI-friendly reports, SARIF output, ATT&CK-style mapping, and test artifacts that run the same way twice.
|
Local MCP / AI attack-surface audit and evidence gate for authorized testing. SDK, CLI, MCP, Docker sandboxing, and CI artifacts. |
Scans GitHub and local git for leaked AI API keys across 50+ platforms. Pairs each key with its relay URL and checks whether it still works. |
|
Multi-platform agent security testing with DAG-style attack chains and Rust components. Covers 9 agent platforms with exploit modules and ATT&CK mapping. |
Evolutionary framework for LLM security assessment: 6 attack methods, 18 model adapters, and a Vue 3 dashboard for structured resilience measurement. |
|
MCP server for CTF workflows: 126 tools across web, crypto, and pwn, with risk levels and warnings on the dangerous ones. |
Installable Codex / agent-skills workflow pack for dev, Agent / RAG, API / data, Office artifacts, research diagrams, and authorized security review. |
- Authorized testing only. Research code is for owned systems, lab environments, CTFs, or explicitly approved assessments.
- Evidence over claims. Reproducible commands, logs, SARIF, screenshots, and minimal proof artifacts — not assertions.
- Defender-first. Methods are published in a way that helps detection, hardening, and regression testing.
- Small tools, sharp boundaries. No hidden production assumptions, no credential leakage, no irreversible actions.
Private contribution visibility is controlled by GitHub profile settings. The graph below is generated from visible profile activity instead of third-party stats cards.
COFF0XC :: break carefully, document clearly




