AI is accelerating vulnerability discovery. The harder question is whether organizations can turn those findings into deployed fixes quickly enough. On October 21, PHP released emergency fixes for two critical vulnerabilities in its core engine. Earlier this month, Microsoft’s Patch Tuesday and AI-assisted discovery brought another record volume of disclosures. AI-assisted discovery is helping drive this pressure. Finding a bug is only the beginning of the work.

The same week, Sam Altman said he no longer thinks we can just patch all the bugs. That is the right question to put to infrastructure teams: if discovery outruns patching, what changes? This is part one of a pair. Part two, From Cloudburst to Containment, has Altman’s full remarks and covers the other side: what happens when the AI itself is the thing trying to get out.

The new math of Patch Tuesday

Security teams built their processes around a predictable monthly patch cycle. Recent disclosure volumes are straining that rhythm, and the work extends well beyond reading advisories: teams must identify affected systems, assess exposure, test updates and deploy them safely.

The tracker series below illustrates the change over the twelve Patch Tuesdays from November 2025 through October 2026, compared with the same months two years earlier. The reported totals are roughly 4.5 times higher. Because trackers differ in scope and counting conventions, treat that ratio as an indication of scale rather than a precise measure of growth in unique Microsoft vulnerabilities.

Figure 1 — Reported Microsoft CVEs per Patch Tuesday, November 2025–October 2026 versus November 2023–October 2024. Compiled tracker counts differ in scope; December 2023 is approximate. The series illustrates the trend and is not a normalized count of unique vulnerabilities.

.

The sharpest rise comes in the last five months. AI can help researchers inspect more code and investigate more suspected flaws, but disclosure totals alone cannot establish how much of the increase AI caused. For infrastructure teams, the immediate consequence is clear: more findings compete for the same testing and deployment capacity.

Figure 2 — Reported Patch Tuesday totals through October 2026 compared with earlier full years. Differences in tracker scope limit precise year-to-year comparisons.

PHP2Shell shows the stakes

PHP2Shell makes the operational risk concrete. PHP’s October 21 advisory identifies CVE-2026-109618 and CVE-2026-109619 as critical vulnerabilities that can permit unauthenticated remote code execution. They involve memory handling during request processing, including out-of-memory failures and stale internal state. Because the flaws are in PHP itself, exposure extends across applications built on the engine. Source

Public patches also give attackers a map of what changed. AI can assist that analysis, increasing the pressure to close the gap between disclosure and deployment. The practical question is how long an exposed service remains reachable while its update moves through the organization.

The same week, the blast radius widened again. Aikido Security reported on October 23 that CVE-2026-111497, previously associated with Node.js (node2shell) and Python, had evolved into what the firm calls HTTP2Shell. The issue affects many libraries and applications that handle HTTP traffic: pre-authentication remote code execution via Brotli decompression, with a dependency footprint that includes extremely common packages. Aikido warned that fixes were not yet available for every affected service and advised teams to watch for Brotli 1.2.1 upgrades as maintainers ship them. Aikido Security on X

The same overload is also driving new triage tools. On October 22, Sherwood Callaway launched cve.now, a free feed that aggregates four vulnerability sources and uses AI to explain each item in plain English, with no login required. His launch post pointed at the practical problem behind the headlines: php2shell, LiteLLM, Railway, and other incidents landing in the same news cycle. Callaway on X

AI as the bug hunter

The capabilities extend beyond individual discoveries. Systems such as Google’s Big Sleep investigate vulnerabilities, while DARPA’s AI Cyber Challenge has explored automated discovery and repair. These efforts matter because they connect finding a flaw with demonstrating its impact and proposing a fix. The operational value depends on the quality of that evidence and the work needed to act on it.

Every finding still needs judgment

VulnCheck’s State of Exploitation report for the first half of 2026 offers a useful counterweight to the headlines. Of 1,061 vulnerabilities attributed to AI-assisted discovery, 14 (1.3%) had confirmed exploitation in the wild, roughly matching the overall rate for that period. Confirmed exploitation is an observed minimum, and recently disclosed flaws have had less time to accumulate evidence. Even so, discovery volume should not be confused with demonstrated attack activity. Source

Figure 3 — AI findings, published CVEs and confirmed exploitation are different measures. Source: VulnCheck State of Exploitation 1H 2026. These populations are not a single tracked cohort or a conversion funnel.

There is also a cost to weak reports. In January 2026, curl’s Daniel Stenberg announced the end of monetary bug bounties and a move away from HackerOne as the recommended reporting channel, describing the burden of fabricated and poorly researched submissions. Maintainer time spent disproving a confident report is time unavailable for investigating a real flaw. Source

The challenge is to handle both real discoveries and low-quality submissions without giving them equal weight. A growing disclosure volume increases triage work even when observed exploitation rates remain low. Teams need evidence of impact, exposure and exploitability to decide what should move first.

AI as the patch writer

AI is also beginning to reduce the work of repair. Google DeepMind introduced CodeMender in October 2025, reporting 72 security fixes contributed upstream during its first six months. The agent combines model reasoning with analysis and testing tools to investigate causes, propose changes and check for regressions. Its initial workflow included human review before patches went upstream. Source

That is a useful pattern, but a validated source patch still has to become a safe production update. Dependencies need rebuilding, applications need testing, and deployment needs monitoring and a rollback plan. The loop is find, triage, fix, verify and deploy. Measure the time until an exposed system is protected, alongside change failures and service health; a faster patch generator alone does not close that gap.

Figure 4 — The defender’s loop. Measure elapsed time to protection across discovery, triage, repair, validation and deployment, with checks for reliability at each stage.

The consequences reach beyond the server room

Reuters reported on October 23 that the disruption surrounding PHP had reinforced the UK government’s call for households to keep three days of emergency supplies. The broader Be Ready campaign covers interruptions to essential services from several causes. For infrastructure teams, the connection is practical: a software incident can interrupt services people rely on. Resilience planning needs to cover how critical services continue during an outage, who depends on them, and how long restoration can take. Reuters report

What this means for infrastructure teams

Closing the loop

AI is making longstanding vulnerabilities easier to uncover and investigate. It can also help produce and validate repairs. The benefit reaches production only when teams can identify affected systems, make sound risk decisions and deploy fixes safely. The organizations that handle this well will treat patching, verification and recovery as continuous operational work, with clear ownership and a measurable time to protection.

Sources

Chart data in Figures 1 and 2: compiled monthly tracker reports from ZDI, Tenable, Splashtop, Outpost24, Ivanti and ManageEngine. Counts vary by scope, as noted in the captions.

Related on Lots of Data: From Cloudburst to Containment (part two) · Anthropic, Scion, and Git · From VM Sprawl to Agent Sprawl · VAST DataEnclave