~/f4n6 $ grep -r "AI-driven surge in vulnerability disclosures strains coordinated disclosure processes" ./investigations/ --include="*.md"

AI-driven surge in vulnerability disclosures strains coordinated disclosure processes

Jeff Davies 20 Oct 2026 6 min read

1. Executive summary

Cisco Talos reports that Microsoft's monthly Patch Tuesday advisory count has risen from a pre-June average of about 100 to 975 in September 2026 and 1,415 in October, with a further October release of almost 500 more than the prior month. Talos attributes the surge to frontier AI models, which it names Mythos, mining decades-old code in widely deployed libraries and tools, and states that disclosure volume is not translating into exploitation at the same rate: of the 975 September advisories, three saw in-the-wild exploitation, and CISA KEV additions grew 37% year over year against a 105% rise in published CVEs. No new vulnerability class has been found; the models are finding known bug types faster in old code. The risk to EMEA financial services is not the raw count but the compression of patch-to-exploit time and the strain on open-source maintainers who cannot absorb the disclosure rate, which raises the chance that a library in your estate carries a known, unpatched flaw. CVE-2021-44228, the Log4j flaw Talos uses as its historical example, remains CVSS 10.0, in CISA KEV since 2021-12-10, with EPSS at 100%.

2. Regulatory framing

Article Trigger (the fact in this item) Practical impact
DORA Art. 28 Talos reports open-source projects being flooded with AI-generated vulnerability reports and warns that maintainer capacity to patch is uneven across projects Third-party risk assessment must cover open-source components whose maintainers may not ship fixes on your patch cadence; record which critical libraries in your estate depend on small maintainer teams and set a fallback (vendor support, fork, compensating control)
DORA Art. 30 The same maintainer-capacity gap means upstream fix timelines are not contractually guaranteed Where an ICT third-party provider supplies software embedding affected open-source components, contractual provisions should require notification of upstream vulnerabilities and a stated remediation window
NIS2 Art. 21(2)(d) Supply chain exposure through embedded legacy libraries that, as with Log4j, are "not a component that got cataloged in visibility scans" Supply chain security measures must include software composition analysis that inventories transitive and embedded libraries, not just declared dependencies

3. Technical analysis & attack chain

This item is a trend analysis, not a report of a specific intrusion. Talos describes no campaign, no malware and no victim. The attack chain below is the exploitation pattern Talos describes generically, using Log4j as its worked example; it is not a chain observed against a named target in this source.

  1. A vulnerability is disclosed in an entrenched library or tool that is widely deployed but poorly inventoried.
  2. Attackers read the vendor patch to locate the flaw, a practice Talos states is common ("attackers often use vendor patches meant to protect customers to discover where vulnerabilities might exist").
  3. Because the component is predictable and present in many environments, a multi-exploit attack can be deployed "with little variability".
  4. Exploitation succeeds where the component was never catalogued, so defenders cannot confirm exposure quickly.

The Log4j case is the template. CVE-2021-44228 is recorded in NVD as CVSS 10.0, CWE-20 (Improper Input Validation) and CWE-400 (Uncontrolled Resource Consumption), is in CISA KEV since 2021-12-10, and carries an EPSS of 100%. Talos describes the library as ubiquitous and included in nearly every environment, yet not a component that got cataloged in visibility scans, so defenders were "scrambling to find out if they even used the log4j library before starting to concoct a mitigation plan". The mechanism Talos emphasises is scale plus low visibility, not exploit sophistication.

On the disclosure pipeline, Talos reports that vendors under its MAPP agreement share detection guidance or proof-of-concept material for vulnerabilities most likely to be exploited, and that Talos has had to add resources to produce coverage on time for staggered or out-of-band vendor timelines. Larger vendors are coping "for now"; open-source projects are not, and whether a project can absorb a flood of reports depends on whether it has a trusted group to share patch work.

Talos argues the traditional 90-day coordinated disclosure window is under pressure because some practitioners now publish without a pre-disclosure period, on the reasoning that if one AI model can find the flaw, another can, so the vulnerability is effectively already public. Talos states that disclosure guidance is "in flux". It draws a parallel to format string vulnerabilities, discovered as a bug class in 2000, peaking in 2005 and then declining, and suggests AI-assisted discovery of known bug types may follow the same curve once the historical backlog is worked through.

The exploitation figures are the load-bearing part for defenders. Of 975 September Microsoft advisories, three had in-the-wild exploitation. KEV additions rose 37% year over year while published CVEs rose 105%. Talos notes KEV tracks distinct exploited vulnerabilities, not attack volume, and that the same vulnerability is reused against many organisations, so the two figures should not be expected to move in proportion.

Confidence caveats. All figures in this section are single-sourced to Cisco Talos; the CVE-2021-44228 severity, KEV status and EPSS value are corroborated by NVD, CISA KEV and EPSS respectively. Talos does not name the AI model vendor behind Mythos, does not publish the underlying dataset for its Patch Tuesday counts, and does not identify any exploited vulnerability among the three September in-the-wild cases. Treat the trend direction as credible and the specific percentages as vendor-reported until independently reproduced.

4. Mitigation & containment

P1 (within 24 hours)

  • Reconcile your CISA KEV list against installed software this week. CVE-2021-44228 is in KEV, CVSS 10.0, EPSS 100%; if any Log4j-core 2.x below 2.17.1 remains in the estate, treat it as an active exposure regardless of the AI disclosure story.
  • Run a software composition analysis pass that inventories transitive and embedded libraries, not just declared dependencies. Talos's central point is that Log4j was present but uncatalogued; a dependency manifest that misses shaded or bundled copies repeats that failure.
  • For any component you cannot inventory, apply a compensating control (network egress restriction, WAF rule, runtime block) until it is catalogued.

P2 (within 72 hours)

  • Triage the current Microsoft and other vendor advisory backlog by exploitability in your environment, not by CVSS alone. Talos's own figures support this: three exploited out of 975 advisories. Rank by KEV membership first, then EPSS, then whether the component is reachable from an untrusted network.
  • Where a vendor publishes detection content under MAPP or equivalent, ingest it into your detection stack rather than writing rules from the advisory text alone.
  • Identify open-source components in your estate whose upstream maintainer team is small or single-person, and record a fallback plan (commercial support contract, internal fork, or removal) for each.

P3 (within 7 days)

  • Add AI-assisted code review to the build pipeline for new code, using models at least as current as those used to hunt for flaws, as Talos recommends. Run legacy code through the same review in bounded chunks; Talos notes this is cheap for new code and expensive for a large backlog.
  • Set a patch SLA that distinguishes KEV-listed vulnerabilities from the general disclosure stream, so a 1,415-advisory month does not stall the queue.
  • Review coordinated disclosure expectations with your software suppliers: whether they publish without a pre-disclosure window, and what notice you get before a fix ships.

5. Indicators of compromise

No indicators of compromise available in the source material. The item is a trend analysis and names no malware, infrastructure, hash or file artefact. CVE-2021-44228 is referenced as a historical example of an exploited library, not as an indicator observed in a campaign described here.

6. Detection

Insufficient indicators to author detection rules. The source contains no distinctive strings, command-line flags, mutex names, file paths, registry keys or behavioural telemetry from which a YARA or Sigma rule could be built. A rule keyed on the string "log4j" or on CVE identifiers would detect reporting about the flaw, not the flaw itself.

CVE assessment

1 referenced CVE — 1 actively exploited (CISA KEV), 1 critical (CVSS ≥ 9.0)

CVE CVSS Exploited EPSS Summary
CVE-2021-44228 10.0 Critical ⚠ KEV 2021-12-10 100% Apache Log4j2 2.0-beta9 through 2.15.0 (excluding security releases 2.12.2, 2.12.3, and 2.3.1) JNDI features used in configurat…

7. Sources

  • Cisco Talos, "There's only so much code: the rise in vulnerability numbers", 2026-10-20. https://blog.talosintelligence.com/theres-only-so-much-code-the-rise-in-vulnerability-numbers/
  • NVD record for CVE-2021-44228 (CVSS 10.0, CWE-20, CWE-400), as supplied in verified reference data.
  • CISA Known Exploited Vulnerabilities catalog, CVE-2021-44228 added 2021-12-10, as supplied in verified reference data.
  • EPSS score for CVE-2021-44228 (100%), as supplied in verified reference data.

8. Adverse Trace position

Adverse Trace assesses this as a strategic trend item with no associated campaign, actor or indicator set, so the operational severity for a given client is low until a specific CVE in the surge is shown to be exploited in their environment. The one hard severity anchor in the item is CVE-2021-44228 at CVSS 10.0, in CISA KEV since 2021-12-10, EPSS 100%, and any client still carrying an unpatched Log4j-core should treat that as the priority over the disclosure-volume story. The Talos figures (975 and 1,415 advisories, three exploited, 37% versus 105% growth) are single-sourced and vendor-reported; we will not repeat them as sector statistics without independent corroboration. Client impact concentrates in two places: patch triage capacity, where a tenfold advisory increase will overwhelm any queue ranked by CVSS alone, and open-source supply chain visibility, where the Log4j lesson is that presence without inventory is the failure mode. We will monitor CISA KEV additions against the monthly advisory counts to test whether the exploitation rate is tracking the disclosure rate, and we will flag any named vulnerability from this surge that enters KEV with exploitation observed in financial services.


Read the original source →

Published via PulseTrace — Adverse Trace threat intelligence.

Post this to LinkedIn
Formatting is converted automatically — headings, bullets, a link back & hashtags. Paste straight in.
J
Jeff Davies