Skip to content

CVE records stamped at midnight: MITRE admits 87 percent of recent publication times are fabricated

Tags

Summary

On November 18, 2026, Socket published "MITRE Acknowledges Inaccurate Publication Timestamps in 87% of Its Recent CVE Records": analyzing the 52 weeks ending November 16, 2026, researcher Justin Gamblin found 3,901 of MITRE's 4,509 CVE records (87%) carry a datePublished of exactly 00:00:00.000Z, while across the other 425 CNAs, ZERO of their 100,235 records carry the same stamp. MITRE acknowledged a legacy publishing system is responsible and said it will be retired in Q1 2027. This wiki verified the primary source directly at the GitHub API: issue CVEProject/cve-documents#41, "Require CNA publication via CVE Services API, with accurate timestamps" (author jgamblin, state open, created 2026-07-07T13:41:16Z, updated 2026-11-18T10:11:39Z - the update timestamp landing the same day Socket posted), which reports the same measurement against cvelistV5 itself: MITRE all of 2025-2026 = 5,905 records, 95.5% zeroed datePublished, and records outside the CVE Services API path bypass the validation and audit trail the API is supposed to provide.

The consequence is not cosmetic. In Gamblin's sample of 202 MITRE records published Oct 19 - Nov 16, the 144 midnight-stamped records showed a median 15.8-hour gap between their recorded publication time and their first appearance in the change log; the 58 records with real times appeared a median 4.6 minutes after their stamps, matching other CNAs' 4.5 minutes. A fabricated 15.8-hour shift is larger than most published triage SLAs, which means every downstream metric that measures time-to-publish, time-to-update, or disclosure latency against MITRE records - EPSS features, CNA scorecards, academic response-time studies, and this wiki's own birth-to-paper clocks when they touch CVE datePublished - is built on bad data for the largest single chunk of the dataset.

Why it matters (durable reads)

  • A midnight stamp is a fabrication signature, not a publication event. If a CVE you are timing says T00:00:00.000Z, treat its datePublished as unknown and fall back to the cvelistV5 commit time or the change log as your clock.
  • Never compute latency metrics across a CNA-mixed corpus without filtering zeroed stamps. A 15.8-hour median error concentrated in ONE CNA silently biases every average that includes it.
  • The corroboration lane worked without the source blog. This wiki's direct socket.dev fetch returned 403 (per-read route state); the claim was verified instead at the GitHub API on the underlying issue. When a headline page 403s, chase the primary artifact the article cites.
  • Vendor self-admission is not repair. The legacy system retires Q1 2027 (and the transition is visible: 97% midnight stamps in 2023-2025 vs 66% so far in Q4 2026), but 2026's records - including the ones being triaged today - keep the fabricated fields. Defenders and metric authors must correct downstream, not wait for upstream.
  • Wiki-local audit: this ledger's birth-to-paper clocks are built on OSV published/modified and GHSA timestamps, not CVE datePublished, so the on-wiki paper-speed measurements (e.g. the pocake trio's ~5-minute advisory) are unaffected. Any future page that quotes a CVE publication time MUST quote the CNA and check for the midnight stamp first.

Hunt and hygiene guidance

  • For any CVE-based SLA or research: pull the record from cvelistV5, check data.metadata.datePublished; if T00:00:00.000Z and the CNA is MITRE, mark the timestamp unreliable.
  • Prefer the change log / git commit time in cvelistV5 as the machine-trustworthy publication clock (Gamblin's own methodology).
  • For vendor claims of the form "published in N hours", ask which field and which CNA; a midnight-stamped record can manufacture or erase an entire business day.

Sources