Inspiration

  • Security teams face too many vulnerabilities.
  • Analysts cannot investigate everything at once.
  • Severity scores alone do not show which vulnerabilities are most likely to be exploited.
  • SonarScan helps identify what should be reviewed first.

What it does

  • Ranks vulnerabilities by risk.
  • Combines exploitation probability, public exploit evidence, historical data, and vulnerability descriptions.
  • Checks evidence from:
    • Exploit Database
    • Metasploit
    • Nuclei
    • GitHub proof-of-concept projects
  • Shows results in a prioritized top-500 list.

How we built it

  • FastAPI backend
  • React frontend
  • IBM Granite embedding model
  • Logistic Regression risk model
  • Known Exploited Vulnerabilities data
  • Monthly backtesting using 2024 and 2025 data

The exploit-signal score is:

$$ \text{Exploit-signal score} = \frac{\text{Positive exploit signals}}{4} $$

The direct-risk score is:

$$ \text{Direct-risk score} = 0.6 \times \text{Exploitation probability} + 0.4 \times \text{Exploit-signal score} $$

Challenges we ran into

  • Combining data from different exploit sources.
  • Matching vulnerability identifiers correctly.
  • Avoiding future information during training.
  • Explaining machine-learning results clearly.
  • Evaluating rare exploitation events.

Accomplishments that we're proud of

  • Built a complete vulnerability prioritization workflow.
  • Combined machine learning with real exploit evidence.
  • Created a four-source exploit-signal score.
  • Added historical ranking evaluation.
  • Built views for queues, rankings, signals, and recommendations.
  • Found that combined exploit evidence performed better than basic severity scoring.

What we learned

  • Severity alone is not enough.
  • Multiple evidence sources create stronger rankings.
  • Historical testing must respect the timeline.
  • Machine-learning scores need clear explanations.
  • Security tools must help analysts make decisions quickly.

What's next for SonarScan

  • Test the IBM Granite model directly in monthly backtesting.
  • Improve score calibration.
  • Add newer exploitation data.
  • Compare the top 50, top 100, and top 500 results.
  • Add asset exposure and business importance.
  • Improve recommendations.
  • Make every ranking decision traceable to its evidence.

SonarScan does not predict every future attack.
It helps defenders decide what deserves attention now.

Built With

  • cvss
  • cybersecurity
  • embeddings
  • epss
  • fastapi
  • hugging-face
  • ibm-granite
  • kev
  • logistic-regression
  • numpy
  • pandas
  • python
  • react
  • recharts
  • security-operations
  • sklearn
  • tailwind
  • threat-intelligence
  • uvicorn
  • vite
  • vulnerability-prioritization
  • vunerability-management
Share this project:

Updates

Submission history