Inspiration

More than 92,000 vulnerabilities were published in the first nine months of 2026, and fewer than 1 in 250 will ever be confirmed as exploited. Security teams can't patch them all, so most sort by CVSS severity. Severity tells you how bad an attack could be, not whether anyone will bother. When we tried that approach on 2025 data, sorting by CVSS put only 7 later-exploited vulnerabilities in the top 500. We wanted to answer the question IBM hears from clients every day, "what do we fix first?", with evidence instead of severity alone, and to show the engineer why.

What it does

GraniteGuard is a triage workspace for vulnerability teams.

  • Ranks what to fix first. A leakage-free ensemble model ranks every 2026 scoreboard CVE by how likely it is to be added to CISA's Known Exploited Vulnerabilities list, and produces our top500.csv.
  • Explains each pick. IBM Granite writes a plain-language "why fix this first" brief for each of the top 500, covering attack surface, weaponization risk and mitigation, with IBM X-Force and reference links.
  • Answers questions. An "Ask Copilot" panel uses Granite plus vector search over the CVE cards in Snowflake to answer questions like "which of these affect edge devices?", and it cites the CVEs it drew on.
  • Shows the trend. Live time-series charts from Tiger Data (TimescaleDB) show how exploit activity is moving, overall and per vendor.
  • Tracks the work. Every vulnerability carries a fix status (Not handled, Fixed, Mitigated, Not affected, Risk accepted). When an engineer signs off a fix with their own wallet, a SHA-256 fingerprint of the sign-off is anchored on Solana. That gives auditors a tamper-proof "Proof of Patch" without putting any CVE, product or company name on a public ledger.

How we built it

Data and time series (Tiger Data / PostgreSQL). We loaded the challenge data into Postgres and stored the 6.4 million monthly exploit-signal rows in a TimescaleDB hypertable with continuous aggregates. SQL window functions build "as-of" features: for each CVE and each snapshot month, only what was known at the end of that month. Examples are how many of this assigner's or vendor's CVEs were already on KEV, and the CVE's EPSS percentile within its month.

A training set that mirrors the test. We replayed 2024 and 2025 month by month under the exact scoreboard rules:

  • The snapshot is the end of the publish month (July for June CVEs).
  • CVEs already on KEV at the snapshot are dropped.
  • The label is "added to KEV later", within a horizon that never looks past what was known.

That gave 82,424 training rows with 112 positives. We deliberately excluded the columns the challenge flags as leaky: nvd_status, n_affected_cpes, n_references, blank scores, and post-attack news or threat-report references.

The model. We used a rank-averaged ensemble of four rankers:

  • LightGBM with monotone constraints, so more exploit evidence can never lower a score
  • XGBoost
  • Logistic regression
  • The "public exploit, then EPSS" baseline

Inputs are EPSS, public exploit flags (GitHub PoC, Nuclei, Exploit-DB, Metasploit), a TF-IDF text prior on descriptions refit each January, assigner, vendor and product KEV history, and time-to-horizon.

Validation. We trained on 2024 and tested on 2025 CVEs under the same rules the judges use for 2026, then confirmed the result on five rolling quarterly folds.

IBM Granite. We used granite-embedding-30m-english to embed CVE descriptions, which power the Copilot's Snowflake vector search, and granite-3.3-2b-instruct to write the threat briefs and Copilot answers. Granite is served from a GPU behind a small FastAPI service.

App. The dashboard is a Streamlit app with Plotly charts. The Snowflake SQL REST API uses key-pair auth and a read-only role. The Solana devnet transactions are built byte by byte with Ed25519 signing, and engineers sign them in Phantom or Solflare.

Challenges we ran into

  • Leakage everywhere. The NIST files are today's versions, and NIST reviews attacked CVEs first, so "has a product list" or "has many references" quietly predicts the answer. Building strict as-of features and auditing every column took longer than training the models.
  • Extreme class imbalance. In the 2025 backtest, only 45 of 35,277 candidate CVEs were later added to KEV. Small changes look like big wins by chance, so we judged every idea on rolling folds rather than a single split.
  • Honest negative results. We tested about 25 improvement ideas, including Granite embedding features, older CVE history, data cleaning and alternative models. None beat the ensemble outside the noise, so we kept the simpler model and documented the results.
  • Infrastructure. We hit blocked package mirrors and model downloads, long embedding jobs that needed to become resumable, and a GPU-hosted Copilot that had to be reachable from the dashboard securely.

Accomplishments that we're proud of

  • On a held-out 2025 test run under the 2026 rules, our ranking found 20 later-exploited vulnerabilities in its top 500. The challenge baselines found 7 (CVSS first), 8 (EPSS first) and 16 (public exploit, then EPSS). Scaled to the 2025 candidate pool, it found 17 against 7, 8 and 14.
  • The strongest signals were ones CVSS and EPSS don't use: what the description says about the product and bug class, and the track record of the assigning vendor's security team.
  • The pipeline is cutoff-safe end to end, and we wrote down exactly which columns we used and why.
  • Every pick is explainable. An engineer sees the score, the signals behind it and a Granite-written brief, not just a number.
  • Proof of Patch adds accountability without leaking a company's patch status to attackers.

What we learned

  • Severity and likelihood are different questions. The best predictor combined "is there public exploit code" with "is this the kind of product attackers go after".
  • In security data, when you knew something matters as much as what you knew. A leakage-free backtest is what makes a ranking trustworthy.
  • A smaller, better-validated model beats a clever one you can't defend. Several ideas that looked promising on one split disappeared on rolling folds.
  • LLMs are most useful here as explainers and research assistants, not as the ranking itself.

What's next for GraniteGuard

  • Add an organization's own asset inventory and exposure, so the queue reflects "exploitable and we run it".
  • Run Granite TimeSeries (TTM) on the Tiger Data EPSS history to forecast which CVEs are heating up before they are exploited.
  • Retrain automatically each month as new KEV entries arrive, and track drift.
  • Ground mitigation advice in vendor advisories, with verified citations.
  • Integrate with ticketing (Jira, ServiceNow) and IBM Concert Protect, so a ranked CVE becomes an assigned fix with a Proof of Patch at the end.

Built With

Share this project:

Updates

Submission history