The AI Paradox in Vulnerability Ecosystem - Workshop At ENISA

GCVE
1

The AI Paradox in Vulnerability Ecosystem

CIRCL approaches
Alexandre Dulaunoy — info@circl.lu
ENISA Workshop 2026
CIRCL — https://www.circl.lu
Initiatives: https://gcve.eu/ · https://vulnerability-lookup.org

The AI Paradox

The paradox

AI makes vulnerability discovery faster and cheaper.

Validation, coordination and publication still require scarce human expertise.

Key message: The bottleneck shifts from finding vulnerabilities to transforming findings into trustworthy vulnerability intelligence.

Operational response: assist, trace, keep control

Assist Trace Keep control
Analyse patches Evidence and sources Local, open tooling
Suggest CVSS/CWE Models and assumptions Deterministic checks
Draft records Human review state Human publication

AI contributes annotations and drafts; the CNA, GNA or CSIRT remains accountable for disclosure.

Provenance by Design

GCVE BCP extensions: make assistance auditable

Extension Question it answers
BCP-05-X-01 Who used AI, for which fields, with what model and human review?
BCP-05-X-02 Which patch and assumptions produced the draft? What supports the CVSS/CWE/CAPEC conclusions?
BCP-05-X-03 Which known dates form the vulnerability timeline, including patch/commit evidence?

Machine-readable provenance helps analysts review AI output without treating the model as an authority.

Reference: Index of /bcp/extension

From Patches to Reviewed Records

patch2vuln: from a fix to a structured draft

flowchart LR
    A["Patch evidence<br/>Git commits / URLs"] --> B["Deterministic checks<br/>Parse, validate, hash"]
    B --> C["Local Ollama<br/>AI-assisted analysis"]
    C --> D["Structured draft<br/>CVE / GCVE + more"]
    D -.-> E["Evidence travels with the draft<br/>Patch hashes, model details, assumptions,<br/>suggested classifications, review state"]

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    classDef evidence fill:#f5f8fc,stroke:#98aeca,color:#23445c,stroke-dasharray:4 3;
    class A source;
    class B,C compute;
    class D output;
    class E evidence;
  • Proposes affected products, descriptions, CWE/CAPEC, CVSS v4, SSVC and remediation.
  • Calculates CVSS scores deterministically from vectors; keeps proposals reviewable.

Repository: GitHub - gcve-eu/gcve-lab-patch2vuln: patch2vuln is a command-line tool that turns a git-format patch into a structured draft vulnerability advisory using a locally hosted Ollama model. · GitHub

Recent patch2vuln improvements

  • Better evidence: Multiple related commits, release-tag checks for fixed versions and CPE 2.3 suggestions reviewed against cpe.gcve.eu.
  • Better analysis: Multi-model comparison, disagreement flags, CVSS v4 and SSVC v2.1 Supplier decisions.
  • Better provenance: AI annotation, patch-to-vulnerability evidence and a patch-based timeline (BCP-05-X-01/02/03).
  • Better access: Web interface, REST/OpenAPI and asynchronous analysis jobs (v1.3.0).

A reproducible, inspectable starting point for analysts—not an automatically approved advisory.

patch2vuln v1.3: queued analysis service

flowchart LR
    A["Web UI / REST clients"] --> B["Flask API<br/>Submit / poll"]
    B --> C[("SQLite queue<br/>Persist jobs")]
    C --> D["Worker<br/>Embedded or standalone"]
    D --> E["Local Ollama<br/>Run analysis"]
    E --> F["Job status + JSON<br/>Downloadable results"]
    F -- "Poll / retrieve" --> A

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    class A,C source;
    class B,D,E compute;
    class F output;
  • Resumable asynchronous analysis, with an embedded or separate worker.
  • Bulk patch inputs, previews, JSON inspection, downloads and OpenAPI documentation.

Documentation: https://github.com/gcve-eu/gcve-lab-patch2vuln/blob/main/docs/web-service.md

Vulniverse integrated into patch2vuln

flowchart LR
    subgraph auto["Automated preparation"]
      direction LR
      A["Patch<br/>Commit / diff"] --> B["CVE 5.2 draft<br/>Generated fields"]
    end
    subgraph analyst["Analyst control"]
      direction LR
      C["Vulniverse<br/>Edit + validate"] --> D["Export JSON<br/>Reviewed draft"]
    end
    B --> C
    D -.-> E["Embedded in the patch2vuln web interface<br/>Schema checks support review;<br/>export does not publish the advisory"]

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef review fill:#fff6e5,stroke:#b77c22,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    classDef evidence fill:#f5f8fc,stroke:#98aeca,color:#23445c,stroke-dasharray:4 3;
    class A source;
    class B compute;
    class C review;
    class D output;
    class E evidence;
  • Inspect the AI output, correct individual fields and validate the CVE schema.
  • Other generated advisory formats remain available for download.

Repository: GitHub - vulnerability-lookup/vulniverse: Vulniverse is a web-based editor for authoring, reviewing, validating, and storing vulnerability records. It is intended as a modern replacement for Vulnogram, with integration as a first-class goal: use the complete standalone application, embed the editor in Vulnerability-Lookup, or connect it to any software that works with the CVE Record Format · GitHub

Vulnerability-Lookup: an integrated analyst workflow

flowchart TB
    subgraph VL["Vulnerability-Lookup + embedded Vulniverse"]
      direction LR
      A["Select patch<br/>Analyst request"]
      B["Review proposed fields<br/>Edit + quality checks"] --> C["Save draft<br/>CNA / GNA path"]
    end

    subgraph P2V["patch2vuln service — local inference"]
      direction LR
      D["API proxy<br/>Submit job"] --> E["Job queue + worker<br/>Local Ollama"]
      E --> F["Draft + provenance<br/>Suggested fields"]
    end

    A --> D
    F -- "Result to review" --> B

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef review fill:#fff6e5,stroke:#b77c22,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    class A source;
    class D,E,F compute;
    class B review;
    class C output;

The analyst chooses which suggestions enter the record; publication remains a separate, authorised CVD step.

Documentation: AI-Assisted Vulnerability Records — Vulnerability-Lookup

The human review is an explicit boundary

Automate Validate before publication
Patch parsing and summarisation Vulnerability and exploitability
Draft classifications and metrics Affected versions / products
Format conversion and provenance CVSS, CWE, CAPEC, SSVC
Queueing and repeatable checks Credits, coordination, disclosure

Review boundary: Model output is a draft; security judgement, publication authority and accountability stay with the analyst and disclosure coordinator.

Interoperability: one analysis, several formats

flowchart TB
    A["Patch evidence + AI analysis<br/>One reviewable source"]
    A --> B["Shared provenance<br/>Evidence, model, assumptions, human review"]
    B --> C["CVE 5.2<br/>Draft"]
    B --> D["GCVE BCP-05<br/>Draft"]
    B --> E["CSAF 2.1<br/>Draft"]
    B --> F["OSV<br/>Draft"]

    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef evidence fill:#f5f8fc,stroke:#98aeca,color:#23445c,stroke-dasharray:4 3;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    class A compute;
    class B evidence;
    class C,D,E,F output;
  • Different publication formats should not lose the supporting evidence and rationale.
  • Vulnerability-Lookup connects editorial work to CNA/GNA publication services.

Interoperability is more than JSON conversion: review and provenance must survive.

The CLI is still useful

python3 patch2vuln.py \
  --gcve-id GCVE-1-2026-20206 \
  --vendor misp --product misp \
  --all-formats-output advisories \
  https://github.com/MISP/misp/commit/382188d2f.patch
  • Produces CVE, GCVE, CSAF and OSV drafts for analyst review before publication.
  • CLI, web interface and API share the same analysis engine.
  • Analyst-supplied facts take priority over model inference.

Scalable Analysis and CVD

Complementary enrichment: VLAI and AI annotations

flowchart TB
    subgraph training["Model development"]
      direction LR
      A["Historical data<br/>Vulnerability-Lookup"] --> B["Fine-tune<br/>RoBERTa"]
      B --> C["VLAI model<br/>Local weights"]
    end
    subgraph inference["Local inference / enrichment"]
      direction LR
      D["New description<br/>Incoming record"] --> E["VLAI inference<br/>Severity class"]
      E --> F["Annotation<br/>Vulnerability-Lookup"]
    end
    C -- "Model weights" --> E

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    class A,D source;
    class B,C,E compute;
    class F output;
  • Early severity suggestion when authoritative scores are not yet available.
  • gcve-eu-ai-extension can add BCP-05-X-01 provenance to AI annotations.

Documentation: AI datasets and VLAI models – Vulnerability-Lookup

From increased discovery to scalable disclosure

flowchart LR
    A["DISCOVER<br/>Patches / reports<br/>Vulnerability evidence"] --> B["ASSIST<br/>patch2vuln<br/>Draft + provenance"]
    B --> C["VALIDATE<br/>Human review<br/>Vulniverse + VL"]
    C --> D["PUBLISH<br/>CNA / GNA<br/>Publication"]
    D --> E["After publication<br/>Sightings, enrichment,<br/>correlation, record updates"]
    E -- "Review updates" --> C

    classDef source fill:#edf4fc,stroke:#2668a0,color:#23445c,stroke-width:1px;
    classDef compute fill:#eaf8f6,stroke:#168c91,color:#23445c,stroke-width:1px;
    classDef review fill:#fff6e5,stroke:#b77c22,color:#23445c,stroke-width:1px;
    classDef output fill:#eef0f3,stroke:#23445c,color:#23445c,stroke-width:1px;
    classDef evidence fill:#f5f8fc,stroke:#98aeca,color:#23445c,stroke-dasharray:4 3;
    class A source;
    class B compute;
    class C review;
    class D output;
    class E evidence;

Local AI increases analyst capacity; controlled disclosure and open standards preserve trust.

Takeaways

  1. The bottleneck has moved: More discoveries mean more validation and CVD work.
  2. Provenance matters: GCVE BCP-05-X-01/02/03 expose assistance, evidence and timelines.
  3. patch2vuln now scales: Queued web/API jobs, bulk drafts, richer classifications and multiple output formats.
  4. The integrations close the loop: Vulniverse for review; Vulnerability-Lookup for coordinated recording and publication.
  5. Humans decide: Automation drafts and enriches, analysts verify and publish.

Thank you for your attention

:link: PDF Presentation (203.9 KB)