NoiseGate

Filter the noise from bug bounty reports.

A program-aware AI triage assistant that evaluates report quality, scope alignment, and impact — so researchers submit better reports and triagers focus on real vulnerabilities.

Features

• Three-verdict system — PASS / WARN / FAIL maps directly to downstream actions for both researchers and triagers
• Program-aware — import any bug bounty program from a URL or local file; NoiseGate extracts scope, ineligible findings, and rules using an LLM
• Structured output — every verdict includes per-check reasoning (scope, eligibility, quality, impact)
• Multi-provider — Anthropic, OpenAI, Gemini, or local Ollama via a single --model flag
• Scriptable — --json output + exit codes (0=PASS, 1=WARN, 2=FAIL, 3=error) for automation and pipelines
• No PoC, no FAIL — a real finding with a weak writeup gets WARN, not FAIL; the finding is preserved

Verdicts

Verdict	Meaning	Researcher action	Triager action
PASS	In scope, real impact, clear writeup	Submit it	Triage it
WARN	Real finding, incomplete writeup	Fix the report	Ask for more info
FAIL	Out of scope, no real impact, non-qualifying	Don’t submit	Close it

Installation

Requires Python 3.12+.

pip install noisegate                    # Anthropic (default)
pip install "noisegate[openai]"          # + OpenAI
pip install "noisegate[google]"          # + Gemini

Or from source:

git clone https://github.com/sgmurphy/noisegate.git
cd noisegate
pip install -e ".[dev]"
playwright install chromium

Quickstart

1. Configure your LLM provider:

noisegate llm

2. Import a bug bounty program:

noisegate program add acme --url https://hackerone.com/acme

3. Triage a report against the program:

noisegate triage report.md

Usage

Triage a report

# Uses the default program
noisegate triage report.md

# Override the program for this run
noisegate triage report.md --program acme

# Machine-readable JSON (for CI/scripts)
noisegate triage report.md --json

# Use a specific model
noisegate triage report.md --model openai:gpt-4o
noisegate triage report.md --model google-gla:gemini-2.0-flash
noisegate triage report.md --model ollama:llama3

Exit codes: 0=PASS, 1=WARN, 2=FAIL, 3=error

Multi-provider support

NoiseGate uses pydantic-ai for LLM abstraction. Pass --model <provider>:<model-id> to switch providers:

Provider	Model string	API key env var
Anthropic (default)	anthropic:claude-sonnet-4-6	ANTHROPIC_API_KEY
OpenAI	openai:gpt-4o	OPENAI_API_KEY
Gemini	google-gla:gemini-2.0-flash	GOOGLE_API_KEY
Ollama (local)	ollama:llama3	(none)

JSON output

{
  "checks": [
    {
      "check_id": "asset_in_scope",
      "passed": true,
      "reason": "The reported target matches the in-scope asset *.example.com."
    },
    {
      "check_id": "finding_ineligible",
      "passed": true,
      "reason": "The finding type is not listed as ineligible."
    },
    {
      "check_id": "required_sections_present",
      "passed": false,
      "reason": "No proof-of-concept is included, which the program requires."
    }
  ],
  "summary": "The report targets an in-scope asset with real impact, but is missing a required proof-of-concept.",
  "verdict": "WARN",
  "model_used": "anthropic:claude-sonnet-4-6",
  "prompt_tokens": 1234,
  "completion_tokens": 567
}

This site is open source. Improve this page.