Inspiration
Bug-bounty hunters lose time and rewards to preventable triage rejections: missing proof-of-concept, severity that does not match the impact, steps that cannot be reproduced, reports too vague to act on. Triagers see the same mistakes every day — but hunters get that feedback only after submitting, days later, if at all. Bounty Triage Copilot gives that feedback in under a second, before submission. Paste a draft report, pick the program, and the app scores it 0-100 against a four-dimension rubric — PoC evidence, severity justification, reproducibility, and clarity — cross-checks the claimed severity against the impact described, screens for common duplicate patterns, and returns concrete, actionable suggestions. Everything runs locally: the scoring engine is a deterministic rules-based system in pure Python (standard library only, no dependencies), and the optional second opinion comes from a local LLM (Ollama, qwen2.5:3b) on localhost. No accounts, no uploads, no data leaves the machine. If the LLM is slow or unavailable, the app degrades gracefully to the instant rules-based score instead of failing.
What we built
Web app (Python stdlib HTTP server + plain HTML/CSS/JS, zero dependencies) with live rubric bars, severity cross-check, duplication screen, and suggestions. Rules-based rubric engine (13 unit tests: 8 rubric + 5 LLM-parser) with two built-in example drafts: a strong reentrancy report (scores 93/100, Grade A) and a vague access-control claim (scores 7/100, Grade F). Optional local-LLM review endpoint with strict JSON parsing and graceful fallback. End-to-end smoke test covering every endpoint (9 checks, all passing).
What we learned
A small local model on CPU is honest but slow: a 30-token reply took ~275 s on the demo VM, so the LLM must be an async enhancement, never on the critical path. Designing for graceful degradation mattered more than the model itself. Triage quality is mostly checkable mechanically: presence of a PoC, numbered steps, and severity/impact agreement already separate strong drafts from weak ones without any AI.
What's next
Per-program rubric profiles (severity tables and required evidence differ by program). Real duplicate search against public disclosure databases. A browser-extension 'check before submit' flow and support for faster local models / GPU inference.
How to run
python3 app.py → http://127.0.0.1:8080 (optional LLM: ollama pull qwen2.5:3b && ollama serve)
Built With
- css
- html
- javascript
- ollama
- python