Cisco Talos Ties AI Models to Vulnerability Advisory Surge

Executive Summary

• Cisco Talos reported that monthly Microsoft Patch Tuesday vulnerability advisories have risen more than tenfold, reaching 1,415 in October, against a pre-June average of about 100 per month. Source
• Talos attributes the surge largely to frontier AI models, particularly the introduction of Mythos as of April, mining decades of technical debt in widely used legacy libraries and tools. Source
• Published CVEs rose 105% year over year, while additions to the Known Exploited Vulnerabilities list grew 37%, and only three of the 975 advisories Microsoft issued in September saw in-the-wild exploitation. Source
• Talos says the volume is straining open-source maintainers and pressuring the traditional 90-day coordinated disclosure model, with some practitioners publishing advisories without a disclosure period. Source

Key Takeaways

• Disclosure volume is rising far faster than demonstrated exploitation, based on Talos's comparison of CVE and KEV growth.
• Most of the new findings come from old code in large-profile products, not from newly disclosed vulnerability classes.
• Talos expects discovery to plateau and decline as the historical backlog is remediated and AI checks are embedded in new development pipelines, while labelling this a theoretical path.
• Talos advises defenders to prioritise high-exploitability flaws in their own environments and to build defense-in-depth.

Cisco Talos on the Patch Tuesday Surge

The reported increase is concentrated in a short period. Talos notes that the September Microsoft Patch Tuesday release carried 975 advisories, that October delivered 1,415, and that the pre-June baseline was roughly 100 a month. The authors, Kri Dontje and Yves Younan, place the inflection point around April, when frontier models tuned for vulnerability detection arrived, with the introduction of Mythos as the notable step.

Talos attributes the findings to AI systems working through a backlog of technical debt in libraries and tools that anchor many environments. The report says these models approach code without some of the assumptions human defenders carry, and that they are thorough in enumerating weaknesses. The sources of the findings are repositories of older code behind large-profile products and systems.

Younan is identified in the report as Director of Talos Vulnerability Discovery, Detection, and Testing. Talos says it has needed to invest more resources to deliver coverage on time for critical vendors such as Microsoft, as notifications arrive quickly and on staggered or out-of-band timelines. The report also describes Talos's relationship with large vendors through the Microsoft Active Protections Process agreement, under which organisations share detection guidance or proof-of-concept material for vulnerabilities most likely to be exploited.

Cisco Talos on Disclosure Pressure and the 90-Day Model

The report frames the disclosure system as the second pressure point after raw volume. The established model gives a vendor a set period, typically 90 days, to produce coverage before public release. Talos says that model improved security for decades but is now contested, because some practitioners publish advisories without a notice period on the reasoning that any capable AI could find the same weakness, making the finding effectively public already.

Talos does not resolve that debate. It states that guidance is in flux and that conversations about disclosure expectations are actively changing. The report holds to the position that vulnerabilities should be reported to enable informed defenders and regular patching, and that vendors responsible for vulnerable code need time to create coverage.

The strain is uneven. Talos says larger vendors can handle the pressure for now, while open-source projects are being flooded. Some projects have a trusted group that can share patching work, and others involve the full community, but the report says the choice rests with each project.

Cisco Talos on Why Volume Does Not Equal Exploitation

Talos separates the count of advisories from the count of exploited flaws. Of the 975 vulnerabilities Microsoft reported in September, the report says three in-the-wild exploits were detected. On a year-over-year basis, KEV additions rose 37% against a 105% rise in published CVEs. Talos states that these figures would not necessarily move in proportion, because adversaries need only enough useful vulnerabilities to compromise targets and can reuse the same flaws across many organisations. The KEV list tracks distinct vulnerabilities known to be exploited, not attack volume or successful compromises.

The report also notes that many vulnerabilities are impractical for attackers to use, and that AI can ingest code, advisories and attack surfaces at speed but cannot currently judge the value of a weakness. On the attacker side, Talos says AI can remove the patch-parsing bottleneck that has slowed human attackers, and that awareness of vulnerabilities often leads to more exploitation.

Cisco Talos on Whether Discovery Will Plateau

The geography column is left unfilled because the source does not provide geographic detail for any row.

Talos looks to the format string vulnerability class, first identified in 2000, as a precedent. Discoveries ramped to a peak in 2005 and then declined over the following years, and the class remains present but is a smaller share of annual CVEs. The report says frontier models have not found a new vulnerability class; they are more efficient at finding known types. Because essential software has a long release history, there is plenty of code left to mine, but Talos argues there is only so much code.

The expectation offered is that the historical backlog will be turned over and patched, and that new code will run through AI bug-checkers, allowing the rate of new vulnerabilities to level off and fall. Talos labels this a theoretical path and attaches caveats, including the risk that more awareness drives more exploitation and that AI accelerates attacker patch analysis.

Cisco Talos on Defensive Priorities

Talos describes the core danger as exploitation time falling rapidly while the pool of vulnerabilities that could become exploits is still rising. It says the main difficulty for defenders is judging which vulnerabilities are real for their environment. Its recommendations are differentiated by role. Teams writing code should add AI evaluation by models at least as current as those used by hunters, run new code through cheap checks, and work through legacy backlogs in chunks, which the report describes as costly to process in full.

For teams keeping tools ready, Talos advises upgrading as soon as possible and going further: for Cisco customers, using LiveProtect to mitigate between advisory release and patching, moving beyond static domain blocking to behavioural detection that alerts on anomalous traffic toward automation platforms, and concentrating on the highest-exploitability flaws in their own environments. The report also recommends defense-in-depth through multiple protection systems, custom IPS/Snort triggers using SnortML, and interim mitigations to cover patching delays.

Cisco Implementation Risks

Editorial independence: this analysis was prepared independently from the public source material and reflects no input from Cisco or Cisco Talos. Source note: Cisco Talos report.

What This Means for Practitioners

For security and engineering leaders, the report shifts the problem from tracking advisories to deciding which ones matter. The gap between CVE growth and KEV growth suggests that triage capacity should go to flaws with a viable path into the organisation's own environment, rather than to the total count. Talos's recommended levers, AI checks in build pipelines, staged remediation of legacy code, interim mitigations and behavioural detection, each consume budget and staff time, so sequencing matters. The highest-value early step is likely a defensible inventory of exploitable exposure, because the report says open-source projects are flooded and vendor coverage arrives on staggered timelines.