Security Research Manager at CyCraft Technology, based in Taiwan. I maintain awesome-web-security — a curated web-security learning list, 13.6k★ / 1,800+ forks — and lately I've been building infrastructure for LLM agents that have to get security judgment right, not just look plausible.
Four agents, each pointed at a real problem instead of a benchmark. Two of them are things you can install and use today:
diff-sentry — catch a malicious pull request before you
merge it. A GitHub Action that reads a change the way an attacker hopes you won't: pull_request_target
that checks out the PR head, base64 that decodes to a shell payload, secrets reaching a network sink,
payloads shoved past the diff viewport by a whitespace run. Fifteen lines of YAML, and no API key, no
model and no network — so it runs on fork pull requests under the read-only token they already get.
- uses: qazbnm456/diff-sentry@v0.4.0ctx-distillery — your coding agent's memory rots. Distil it. Every project you work on with an AI coding agent accumulates a pile of past conversations too large to read and a memory store that quietly starts contradicting the code. This reads both and proposes what to prune and what is worth promoting into a durable memory file or a reusable Skill — and never writes a byte without your explicit, per-candidate approval. Ships as a PyPI package and as an Agent Skill.
uv tool install "ctx-distillery[cli]"
npx skills add qazbnm456/ctx-distilleryTwo more are research agents rather than products, and honest about it:
| Project | What it does |
|---|---|
| cve-reverser | Reverses a publicly disclosed WordPress CVE from its patch into a local-lab PoC and a Nuclei detection template. Strict planner/lifeline/generator role separation, so no single model both reasons about the bug and writes the final template unchecked. |
| toolscout | Implements Microsoft's ATLAS method: a small planner solves tasks over a huge MCP toolspace by discovering it progressively (list → load → describe → call) instead of holding hundreds of tool schemas in context. |
All four sit on rlm-harness — a reusable harness over DSPy's Recursive Language Model module. Every task gets a full, replayable JSONL trace (main steps, sub-model calls, tool calls), execution is sandboxed by default (pyodide/Deno; raw local execution is refused unless you opt in), and those traces export as reward-free SFT/RL datasets. It is domain-agnostic — security just happens to be my own first use of it.
The through-line: every verdict traces back to evidence rather than being trusted as self-report, every run is exportable as training data, and every README says plainly what is not solved yet. I would rather ship an honest residual-risk section than a demo that only works once.