Skip to content
View qazbnm456's full-sized avatar
:electron:
Focusing
:electron:
Focusing

Highlights

  • Pro

Organizations

@LulumiProject

Block or report qazbnm456

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
qazbnm456/README.md

Hi, I'm Syue Siang Su (Boik Su) 👋

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.


What I'm building right now

A small stack of tools on top of DSPy's Recursive Language Model module, going from general-purpose harness down to real, adversarial security tasks:

rlm-kit — the foundation. A reusable harness over dspy.RLM: every task gets a full, replayable JSONL trace (main steps, sub-model calls, tool calls), sandboxed execution by default (pyodide/Deno; raw local execution is refused unless you opt in), and traces export as reward-free SFT/RL datasets. Domain-agnostic — security just happens to be my own first use of it.

Three consumers built on it, each aimed at a real problem instead of a benchmark:

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 — a traced, trainable agent, not a one-shot script. Strict planner/lifeline/generator role separation so no single model both reasons and writes the final template unchecked.
diff-sentry Reproduces the shape of Datadog's BewAIre defense: classifies a GitHub diff as benign/suspicious/malicious, treating the diff as untrusted data inside a sandboxed REPL, never as instructions. Tested against a reconstruction of the actual hackerbot-claw incident payloads.
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.

The through-line across all three: every verdict is traced back to evidence rather than trusted as self-report, every run is exportable as training data, and every README says plainly what's not solved yet. I'd rather ship an honest residual-risk section than a demo that only works once.


Elsewhere

Blog / writing · X (formerly Twitter) · Patreon

Pinned Loading

  1. rlm-kit rlm-kit Public

    A clean, reusable harness for building tasks on DSPy Recursive Language Models (RLMs).

    Python 4 1

  2. cve-reverser cve-reverser Public

    An autonomous RLM agent that reverses disclosed CVEs from their public patches into local-lab PoCs and Nuclei detection templates.

    Python 2

  3. diff-sentry diff-sentry Public

    An autonomous RLM agent that classifies GitHub changes for malicious intent, reading the diff as untrusted data and emitting evidence-backed SIEM signals.

    Python

  4. toolscout toolscout Public

    An ATLAS-style rollout harness on rlm-kit: a small planner progressively discovers a large MCP toolspace and computes over tool results as code, emitting reward-free trajectories + per-criterion fa…

    Python

  5. awesome-web-security awesome-web-security Public

    🐶 A curated list of Web Security materials and resources.

    Python 13.6k 1.8k