Academic origins · 2023–2025

Built on a master’s thesis.
Not a weekend hackathon.

The detection brain (four signals in the thesis, six in production today) and the XYZ risk score weren’t dreamed up in a sprint planning. The methodology was developed and formally defended as a master’s thesis at the Munster Technological University (formerly Cork Institute of Technology), then refined into the production engine you see today.

  1. 01 Defended at academic committee Formal methodology review at Munster Technological University (formerly Cork Institute of Technology), 2023–2025. Not retrofitted to a marketing site.
  2. 02 Run on real corpora 840K+ deduplicated advisories, 250K+ known-malicious package records and 13M+ dependency edges, all routed through one scoring engine (production database, 12 November 2026).
  3. 03 Validated against history Replays event-stream, ua-parser-js, colors, faker and node-ipc as control cases before any production install touches a developer machine.
[ Methodology paper available on request ] →
The central idea

Phantom critical vulnerabilities,
in both directions.

Jacobs et al. (2020) observed exploitation in the wild for only about 5.5 percent of vulnerabilities. Remediating everything rated CVSS 7 or higher reached about 74 percent of the exploited ones, while roughly nine in ten of the vulnerabilities fixed under that policy were never exploited. The thesis names that gap and defines the class both ways: the loud ones that are not dangerous, and the quiet ones that are.

// type 1 · false alarms

High score, no teeth.

A vulnerability that CVSS calls critical but that nothing in the real world backs up. It is deprioritised so the team stops spending on it.

  • CVSS 8.0 or higher
  • XYZ score below 6.0
  • EPSS below 0.1
  • No public exploit, no exploitation in the wild
9.0 to 5.5worked example, CVSS to XYZ
// type 2 · hidden threats

Low score, real danger.

A vulnerability CVSS rates medium that sits in an attack chain, in a package everything depends on, on a network-exposed path. It is escalated.

  • CVSS below 7.0
  • XYZ score 7.0 or higher
  • Member of a known attack chain
  • High dependency centrality and network exposure
6.5 to 8.2worked example, CVSS to XYZ
// fig. 01 · what each score takes into account
Scored 0 to 2 from the framework comparison in chapter 3: not considered, partial, built in.
Dependency context Attack chains Exploit evidence Data sources AI assessment PCV detection
CVSSEPSSXYZ score
// fig. 02 · same vulnerability, different verdict
Four worked examples of the scoring rules from chapter 5, not measurements of specific CVEs. The score moves down for phantom criticals and up for hidden threats.
0 2 4 6 8 10 High CVSS, no exploit Type 1 PCV · EPSS 0.05 9.0 5.5 Medium CVSS, in attack chain Type 2 PCV · EPSS 0.15 6.5 8.2 High CVSS, active exploits genuine critical · EPSS 0.85 8.5 9.1 Low CVSS, high centrality supply-chain risk · EPSS 0.10 4.0 6.8
CVSSXYZ score
What was measured

The numbers in the thesis,
as of December 2025.

The XYZ score weighs base risk at 40 percent, exploitation history at 30, environmental context at 20 and an AI assessment at 10. It was evaluated on a database built for the thesis and against pip-audit on the same packages. The middle panel repeats that comparison on 12 November 2026.

// database scale, measured
SourceRecords
CVE records717,146
OSV records299,257
Exploit records, ExploitDB46,113
GitHub Security Advisories21,430
Malicious packages4,614
Totalnot comparable*
*The thesis total summed source tables that overlapped, so it is not comparable to a deduplicated count and we no longer quote it. December 2025 snapshot: over 100,000 packages in each of NuGet, npm, Maven and PyPI. On 12 November 2026 the deduplicated database holds 840K+ advisories and 250K+ known-malicious package records.
// re-measured 12 November 2026, same five packages aiohttp 54 55 requests 8 8 urllib3 22 22 certifi 3 3 cryptography 18 18
OSV.devXYZ

Distinct CVEs, OSV.dev vs XYZ105 vs 106

The thesis compared 3 pip-audit findings with 115 XYZ records and presented the gap as a coverage multiple. Re-measured on the deduplicated database with distinct CVEs, XYZ and OSV.dev hold the same CVEs for these packages, so the paper withdraws that claim. The gap was scope (all versions against the installed version) plus duplicate rows.

EPSS score, of the 106 CVEs106

XYZ Score92

NVD names the Python package82

Matching NVD by product name alone would attach six CVEs that belong to other software. Thesis timing (an average 47 seconds for XYZ against 20.4 seconds for pip-audit on 217 packages) was removed in production by precomputing scores; the standalone scanner has not been re-timed.
// attack chains, walked through the model 0 2 4 6 8 10 ProxyShell 6.5 · 7.2 · 6.3 alone 9.8 HAFNIUM 7.3 · 7.8 · 6.5 · 7.8 alone 9.9
Individual CVE, CVSSCombined chain
Bugs rated medium to high that together give remote code execution. CVSS scores each one alone. The combined score is the model’s output for two chains that were already public, a consistency check and not a prediction.
The first production year

What the engine did in production,
read on 12 November 2026.

The paper reports the deployment as measured, including the weeks it caught nothing early. All figures were read from the production database on 12 November 2026.

// install decisions, 6 April to 12 November 2026
Allow536,217
Alert7,121
Block526
Quarantine492
Unverified36
Total544,392
From 1,260 machines: 472,147 npm installs and 71,679 PyPI, with Go and NuGet in the hundreds.
// the June 2026 PyPI worm

First release to first alert31 min

Alert to first public OSV record6 h 28 min

Projects alerted37 of 38

One case, not a rate. It was an in-app alert and nothing was blocked; the first block-list promotions followed on 10 June.
// npm, flagged before OSV listed it / OSV-listed, by week
12 October 20260 / 755
19 October 202665 / 593
26 October 2026271 / 1,049
2 November 202612 / 451
0 to 26 percent a week on npm, and near zero on PyPI. Most npm misses were releases the scanner never opened.
Structure

Six chapters, one argument.

  1. 01IntroductionFalse-positive fatigue, the cost of data fragmentation, and why a score that ignores exploitation cannot prioritise.
  2. 02Literature reviewCVSS and EPSS, the vulnerability databases and their gaps, supply-chain attacks, exploitation prediction, AI in vulnerability assessment, and attack-chain analysis of HAFNIUM, ProxyShell and Log4Shell.
  3. 03DesignEight system objectives, the cross-source integration layer, the dependency analysis framework, the XYZ score, the AI committee, attack-chain detection and the PCV classifier.
  4. 04ImplementationPostgreSQL schema, the local CLI agent, the scoring engine, the daily update pipeline and the API.
  5. 05EvaluationPerformance testing, comparative analysis against pip-audit, and a walkthrough of the ProxyShell and HAFNIUM chains.
  6. 06Conclusions and future workWhich objectives were met, what was not measured, and the roadmap that became the product.
From the thesis to production

The thesis scored known vulnerabilities.
The product extends it to day-zero malware.

Everything below started as a chapter. The one thing the thesis did not attempt, catching a malicious package that has no CVE at the moment of install, is what the same scorer does today at the proxy.

Cross-source database, December 2025 snapshot840K+ advisories, 250K+ known-malicious package records, 13M+ dependency edges, refreshed continuously
XYZ score, 40 / 30 / 20 / 10 weightingSix parallel detection signals fused by a cross-signal scorer into allow, alert, quarantine or block
Local CLI agent, xyz scanxyz Scanner (xyz) on PyPI, the install-time proxy, the VS Code extension and CyberXYZ Gate
Attack-chain detection, ProxyShell and HAFNIUMCampaign tracking mapped to MITRE ATT&CK, and wave correlation across registries
AI committee: designed in the thesis, never evaluatedProduction runs a single model, with code excerpts from the affected release in the prompt, to review alerts

Presented as Beyond CVSS: Detecting Phantom Critical Vulnerabilities in APT Attack Paths at AtlSecCon 2026, Halifax.

// book a demo

Watch it block a real attack. Then run it on your own installs.

Book 15 minutes. We run real packages through CyberXYZ while you watch, including axios 1.14.1, a hijacked release that dropped remote-control malware in March 2026, and show you exactly what your team would see. Then we connect one laptop or one pipeline for a 30-day proof of value.

What happens next

  1. 01A 15-minute callReal packages, real verdicts, and your questions.
  2. 02Connect one laptop or one pipelineOne setup command, or one line in CI. Under an hour.
  3. 03A 30-day proof of value on your own installsEvery signal, no card.

Not ready for a call?

Check any package, free, and see the verdict we would give it.

npm

Try faker@6.6.6ua-parser-js@0.7.29express@4.21.2

Thanks! We'll be in touch.

Check your inbox. We'll reach out within 24 hours.

Where should we send the invite?

  • 15 minutes
  • Video call
  • Times in your time zone

We reply within 24 hours to set a time.

We only use these details to arrange the call. Privacy policy