Gym + code, daily. I build small, honest developer tools and validate before I monetize — evidence over guesses.
Currently building
- agentproof-scan — a
pre-deployment security scanner for self-hosted AI agents. Fires prompt-injection
probes and checks whether the agent leaks its system prompt or API keys — including
in its reasoning trace, where a leak hides even when the answer looks clean. Two-stage
detection (leak vs prompt-disclosure), scan your own HTTP agent or gate CI, plus a
paste-ready handoff block via --handoff.
pip install agentproof-scan· Detect it, prove it.
How I got here
Each tool taught me the limit that pushed me to the next:
- preship — contract checks for FastAPI apps (500 crashes, undocumented status codes, schema mismatches). Learned that contract validation alone can't catch LLM-specific leaks.
- rojaprove — a red-team CLI for LLM apps that detects system-prompt leakage (OWASP LLM07). Learned that watching only the answer channel misses leaks in the reasoning trace — which pointed me to agentproof.
How I work
Solo, local-zero, build-in-public. I'd rather ship a narrow tool that's actually deterministic than a broad one that quietly guesses.
🛠 Python · ⚡ FastAPI · 🤖 LLM / agent security 📌 X @OHS1327
This work follows responsible disclosure principles. The goal is detection, not offense — knowing what leaks before you ship an AI agent, so you can prevent it.
- To prevent misuse, the exact bypass-prompt strings used in testing are masked/generalized.
- All tests run only against intentionally-vulnerable, self-controlled demo targets — never against external services or anyone else's systems.
- What's shared is where defenses shift the leak rate and where they don't, not a runnable recipe for how to attack.