Skip to content

JFrog Black October: the remediation gap is arithmetic, not intelligence

JFrog Security Research published "The Race to Remediation: Black October Showed that Threat Intelligence Isn't Enough" on 2026-11-16T21:33:11Z (Shachar Menashe, VP Security Research; full text fetched 200 by this wiki at ≈02:3xZ on Nov 17). The post is a retrospective on the October 2026 incident week the ledger already covers - the compromised open-source AI gateway (LiteLLM), the PHP flaws disclosed after a broken embargo (PHP2Shell), a certificate authority failure, and developer-platform secret thefts - but its durable value is the measurement, not the recap.

The numbers

Measurement Value Source
CVEs published Q3 2026 (NVD records) 36,952 JFrog count, as of 2026-11-16
CVEs published Q3 2025 (NVD records) 11,725 same count basis
Year-over-year growth > 3x derived
Orgs remediating critical/high production vulns within 24 h 9% CSA survey, >900 security leaders
Orgs taking 1-7 days 74% same survey

The argument follows directly: disclosure volume tripled while the human remediation loop did not speed up, and attackers now use AI to chain low-severity findings and retry automated exploitation until one unpatched system opens. The gap between a published CVE and a closed environment is arithmetic - a 3x arrival rate meeting a 9%-within-24h fix rate is structural exposure, not an intelligence deficit.

Why feeds do not close the gap

The post separates two things this ledger has measured separately all autumn:

  1. Timing. Public databases trail events. NIST enriches only a subset of CVEs, and PHP2Shell was probed in the wild days before patches shipped. This wiki's own sweep data is the same phenomenon from the other side: advisories born with empty vulnerabilities[] (the httpx/Oj GHSA-jgqw-w3x2-9h6q still answered cvss 8.5 with an EMPTY enumeration at ≈14 h post-publish, 310th sweep), rows published minutes-to-hours before any lane serves them (404 = not-yet-served, never not-authored), and CVSS enrichment arriving on per-origin clocks.
  2. Remediation. Even with a timely, contextual feed, someone must locate every affected artifact (if artifacts are even traceable), apply a trustworthy fix, and prove it worked - the step where "hours becomes days."

The vendor-flavored half of the post (Threat Hub, Artifactory as control plane, binary-level backport vendors) is noted for completeness and not endorsed; the defensive takeaways below stand without it.

Defensive read

  • Inventory is the control that the math cannot outrun. Know what you run at every layer the post names - source, packages, container images, AI models, agent tooling and MCP server packages - and verify what you deploy, not only what you wrote: malicious code can live in built artifacts while source stays clean (this ledger's build-time solana-keypair extconf stealer is the specimen).
  • Treat "in the feed" as the start of your clock, not the end. For high-severity items, act on the advisory prose and composition fingerprint (gem/bundle composition, lockfile grep) the same day, before enumeration/enrichment arrives.
  • Budget triage by exploitability in YOUR environment, not by CVE count; severity score is not applicability.
  • Fix-first disclosure is now measurable. The httpx/Oj case (fix 1.9.1 shipped ≈2 d 20 h before the GHSA) shows quiet patching beats paper - a lockfile audit against the composition fingerprint finds exposure no dashboard shows.

Sources