Wired: AI bug flood strains patching as outages hit Microsoft, Apple, NSA

In its weekly Kernel Panic newsletter, Wired's Lily Hay Newman and Matt Burgess address a run of outages over the last week that hit everything from Outlook to Apple, a half dozen US banks, United Airlines and the NSA's classified network. There is no evidence these were cyberattacks, they write, but speculation has run wild, including a feeling that AI is to blame. The authors are blunt: the incidents were definitely not caused by a coordinated AI hacking spree, and taken individually they all seem to have routine explanations.
Their larger argument is that AI is changing the dynamics of digital defense, and that security researchers and analysts have increasingly concluded AI pressures are starting to play a role in real-world tech failures. The chain they describe: AI-assisted vulnerability discovery has unleashed a torrent of software bugs; attackers use AI tools to exploit those bugs faster; that puts more pressure than ever on organizations to patch quickly; and often under-resourced IT teams are left scrambling to evaluate a massive number of patches and push updates to systems that countless people rely on. Researchers and practitioners, they say, keep telling them that AI's biggest impact is that it speeds everything up.
The stated causes are specific for three incidents. Microsoft's 365 outage last week was caused by a flawed update in its infrastructure, exacerbated by other network failures. Officials said this week that the NSA outage, which hit its internal communication system, was due to botched patching following October's critical PHP updates. Apple said its problems this week related to an internal update. US senator Mark Warner said in a statement on Tuesday that AI vulnerability discovery and the resulting patching are 'putting tremendous strain on systems that weren't built to handle that pace of change.' Microsoft CEO Satya Nadella said this week that more 'instability' could be coming, because a 'patch needs to go instantly from being available to being patched in production because it's being exploited in the wild.'
The newsletter adds context. External pressures and bugs in patches are not new, and they have always caused tech failures. Researchers also point to 'tech debt': many organizations worldwide have cut corners for years and failed to invest in upgrading and maintaining their systems. That was always risky, but organizations could gamble with out-of-date systems and typically avoid the worst consequences.
Joe Slowik, director of threat research and cyber engineering at Dataminr, says that even a couple of years ago a vendor would release a patch for a nasty bug and you might have time to validate and apply it. Now that may not even be an option, because turnaround times 'completely defeat the ability to do patch verification.' IT administrators are leaning on AI more to speed up evaluation and prepare systems for updates, but the authors say current AI tools still cannot replace human analysis and planning. Lea Schönherr of the CISPA Helmholtz Center for Information Security, a German national research institution, says AI can help, but people still need to understand what went wrong, check that the patch addresses the cause, and look for similar issues elsewhere. Rolling out patches across widely used systems also takes time. Cloudflare CEO and cofounder Matthew Prince predicts it will feel like the sky is falling for two years, though he suspects AI acceleration will eventually help defenders get on top of common vulnerabilities and better secure software, and that the background noise will 'tick down quite a bit.'
The numbers behind the trend: this week the total number of vulnerabilities found in 2026 passed 100,000. In 2022, the final year before LLMs went mainstream, there were around 25,000 recorded vulnerabilities, and this year's figure, with more than a month to go, is double the combined totals of 2024 and 2025. In the last few days, researchers disclosed using AI to find critical vulnerabilities in PHP (for the second time), the NGINX web server and Java. Forbes also reported 35 vulnerabilities across iPhones, Macs, iPads, Apple Watches and other Apple products discovered by researchers using AI.
The authors flag nuances. Not all AI-discovered vulnerabilities are that impactful. Slowik says the number of vulnerabilities has exploded but the number you need to care about has remained somewhat constant; the caveat is that the pace at which they are weaponized has increased dramatically. Only 300 vulnerabilities have been added to CISA's catalog of exploited vulnerabilities this year. VulnCheck, which runs its own database, counts 1,500 known exploited vulnerabilities so far in 2026, still just 1.5 percent of all bugs discovered this year. Schönherr closes on accountability: responsibility lies with the people and companies that choose to use AI and give it access to infrastructure, and there should be consequences and regulations that can be applied.
The newsletter's In Other News section has three items. JPMorgan Chase, after being affected by the outages, confirmed that in a separate incident attackers accessed certain marketing data held by third parties, which may include customer names and emails; the bank says these security incidents are not linked to any of the outages. Visa and Mastercard also had data affected in attacks on their vendors. Police in the Netherlands said they thwarted several 'attempts at violence' aimed at customers of the digital bank Revolut, which was reportedly compromised in September with customer data stolen, and made several arrests the police say can be linked to the data theft. And a court ruling forced the UK government to admit it ordered Apple to circumvent its own encryption and access customer iCloud data, via a 'technical capability notice' issued last year.
Key facts
- Wired's Kernel Panic says the outages at Outlook, Apple, a half dozen US banks, United Airlines and the NSA were not a coordinated AI hacking spree, and there is no evidence of cyberattacks.
- The authors argue AI-assisted bug discovery and faster exploitation are overloading patching, and that researchers increasingly see AI pressures playing a role in real-world tech failures.
- Stated causes: a flawed infrastructure update plus network failures at Microsoft 365, botched patching after October's PHP updates at the NSA, and an internal update at Apple.
- Vulnerabilities found in 2026 passed 100,000, against around 25,000 in 2022; only 300 are in CISA's exploited-vulnerabilities catalog this year, and VulnCheck counts 1,500 (1.5 percent of the total).
- Cloudflare's Matthew Prince predicts about two years of 'sky is falling' feeling, then calmer conditions as AI helps defenders.
Why it matters
The piece links a visible run of outages to a structural change in security. The source says AI has multiplied discovered vulnerabilities: over 100,000 so far in 2026 versus around 25,000 in 2022, and double the combined 2024 and 2025 totals. Attackers exploit bugs faster, so patches must move from release to production almost at once, as Nadella put it. Slowik says turnaround times now completely defeat the ability to do patch verification.
Who it affects
IT and security teams, often under-resourced, who must evaluate and deploy a huge volume of patches. Also the organizations and people who rely on the affected services: Microsoft 365 and Outlook users, Apple customers, bank customers, United Airlines passengers and the NSA's internal communications. Organizations carrying years of tech debt are named as especially exposed.
How to use it
Read it as a framing for patch management rather than a how-to. The source says IT administrators are using AI to speed evaluation and prepare systems for updates, but that current AI tools cannot replace human analysis and planning. Schönherr says people still need to understand what went wrong, check that a patch fixes the cause and look for similar issues elsewhere. Slowik's distinction helps with triage: the number of vulnerabilities you need to care about has stayed somewhat constant, while the speed of weaponization has risen sharply.
How solid is it
The newsletter's own causes for Microsoft (a flawed update), the NSA (botched patching, per officials) and Apple (an internal update, per Apple) are specific. The AI link is the authors' reading of what security researchers and analysts have concluded, backed by Warner, Nadella, Slowik and Schönherr. No cause is given for the United Airlines or US bank outages. The 100,000 figure is given without a named source. The authors state outright that the incidents were not an AI hacking spree.
Risks and caveats
Not all AI-discovered vulnerabilities matter: only 300 are in CISA's exploited catalog this year, and VulnCheck's 1,500 is still 1.5 percent of the total. External pressure, patch bugs and tech debt are old causes of failures. Prince expects about two years of turbulence before things improve, and Nadella warned of more instability. Schönherr cautions against blaming AI itself, saying responsibility lies with the people and companies that give AI access to infrastructure.
“What matters to me is that we avoid blaming AI when things go wrong”
— Lea Schönherr, tenure-track faculty member at CISPA Helmholtz Center for Information Security