Your Company's Next Knowledge Silo Is Its Coding Agents
Every company has a memory. Some of it lives in code, tickets and runbooks. Some of it has a name, a desk and a badge.
Read articleroot@thinkingtokens:~ $ open research-notes
AI security research Open-source experiments Independent thinking
I’m Nur Gucu. I use a decade of offensive-security experience to study what AI agents trust, how they behave, and how defenders can respond.
// begin here
Three examples of turning a difficult question into something concrete you can read, inspect, or try.
Research into why plain-English instructions can carry software-like risk—and how safe-looking parts can combine into an unsafe system.
Read the DEF CON field report 02 / OPEN SOURCEAn open-source experiment in coordinating complex AI work through reusable skills instead of multiplying agents and infrastructure.
Explore the framework 03 / ANALYSISA hands-on comparison showing what transfers—and what breaks—when the same skill format moves between nine coding assistants.
Open the comparison
currently researching
// behind the terminal
I follow questions until they become something testable: an experiment, a framework, a comparison, or a field note.
The thread running through my work is trust—what agents assume, what systems inherit, and how we can reason about both more clearly.
// research and ideas
For builders, defenders, and curious readers who want the reasoning—not just the conclusion.
Every company has a memory. Some of it lives in code, tickets and runbooks. Some of it has a name, a desk and a badge.
Read articleYour AI agent has a parasite, and it’s written in English. Notes from Red Team Village and Demo Labs, for the people who couldn’t be in the room.
Read articleI built an AI orchestration framework. Then I realized that what I’d built could also serve as a malware delivery system that could be hard to catch.
Read articleHow ORPHEUS reimagines AI orchestration and why Anthropic’s latest thinking confirms the shift is real. GitHub Repository | Core Principles
Read articleAn interactive dashboard comparing SKILL.md implementations across 9 major coding agents — radar charts, spectrum bars, capability scores, and agent datasheets. Companion to...
Read articleIn October 2025, Claude Code shipped SKILL.md — a portable, Markdown-based instruction file that teaches any AI agent how to perform a named task on demand. Within two month...
Read articleLet me tell you about the most dangerous assumption in agentic software development today. It isn’t prompt injection. It isn’t hallucination. It’s quieter than both, and it’...
Read articleThe industry needs to stop treating red teaming as a security gate and start treating it as a continuous practice of failure cartography.
Read articleThe next major class of AI security incidents won’t come from a single prompt injection. They’ll come from interaction topology.
Read article“Vibe coding” became a word almost overnight. Someone said it, it stuck, and suddenly every LLM-powered tool that touches a codebase got slapped with the label. It feels rig...
Read articleNot everything I work on needs a full writeup. Some findings are a single technique, a single vector, a single “oh, that’s interesting” moment. Those live somewhere else — a...
Read articleTraditional offensive security has choreography. You scan, enumerate, find a vector, exploit. Clean. Predictable. There is a playbook — even if it is unwritten — and a decad...
Read articleroot@thinkingtokens:~ $ subscribe
New research, practical field notes, and carefully argued ideas—sent only when there is something worth sharing.