R03 — The Fresh Vulnerability Pipeline

When a new vulnerability lands on Tuesday, how do you get to a reproduced environment, a detection that fires, and a check that proves the fix works — by Thursday, repeatably, without it eating your week?

Most people do this as a heroic weekend and then never again. The point of this volume is to make it a pipeline: a defined sequence with defined inputs and outputs, so the tenth CVE costs you an hour instead of a Saturday.

Why this exists now. The vulnerability data ecosystem changed structurally between 2024 and 2026. The CVE program came within a day of a funding lapse in April 2025. NIST moved to risk-based NVD enrichment in early 2026 — full enrichment now prioritises KEV entries, federal-government software, and EO 14028 critical software, while roughly four out of five incoming CVEs are labelled Not Scheduled with no severity score, no product enumeration, and no enrichment. The landscape fragmented across EUVD and GCVE, and AI-assisted research is a direct cause of the submission surge that broke the old model. The enriched context you used to get for free is now yours to generate.

Six stages — Intake, Triage, Reproduce, Detect, Verify, Publish — each with a defined input and output, so any stage can become a script, a micro-tool, or someone else's job without breaking the chain.

The feed layer: CISA KEV as the highest-signal feed in existence, NVD as the identifier spine you now read knowing most entries are stubs, EUVD as a genuine second opinion, EPSS as the triage multiplier — because a high EPSS score on something not yet in KEV is the pipeline's sweet spot. Plus the vulnerability-intelligence MCP layer and the prompt-injection surface it creates when wired to an agent with write access to your range.

Provenance — why public exploit code is a loaded weapon. Every public PoC is untrusted until you have read it, and reproduced in isolation regardless. If you cannot read it — obfuscated, compiled, or fetching a payload at runtime — that is not a PoC, it is a sample, and it goes under the malware tier rules. Anchored to the 2016–17 leak of nation-state tooling that became two of the most destructive self-propagating malware events in history within weeks of a patch being available: capability weaponises faster than organisations patch, and the only defenders who were ready had already stood it up in a lab.

AI in the pipeline, and the line it does not cross. AI genuinely accelerates triage, reproduction scaffolding, detection engineering and artifact assembly — the highest-value uses, and the ones that directly counter the Not Scheduled problem. Generating novel offensive capability is not a stage in this pipeline and this volume does not provide it. The distinction is not squeamishness: reproducing an already-public vulnerability to build a detection produces a defence; producing a working exploit where no public tooling exists produces the same artifact an attacker wants, whatever the stated intent.

Closes with honest-null outputs ("could not reproduce" and "not detectable with current telemetry" are valid, useful findings), the sightings and info-sharing model, a governance chapter, five worksheets and a readiness checklist.

Includes the CVE Repro Template (ZIP) — the six-part artifact bundle as a copy-per-CVE skeleton: README, pinned docker-compose with no-egress guidance, PoC provenance record, Sigma detection skeleton, remediation check and a logs drop, plus a fully worked example.