AI

Phased patching: how to stop AI-era fixes causing outages

Key takeaways

  • Rushed, estate-wide patching of legacy systems caused outages during a week of digital failures
  • Only 0.45% of 442 Claude-discovered CVEs counted by VulnCheck are being exploited in the wild
  • Prioritise by actual risk and impact, then deploy in phases with monitoring and validation

After reading this, an IT team will be able to build a risk-ranked, phased patching process that closes dangerous holes quickly without taking production systems down with it.

The need is real. Over roughly a week, organisations worldwide saw digital outages that a growing body of industry and government opinion attributes not to AI agents or AI-powered attacks, but to the fear of them. That fear pushed forced upgrades and patches onto decades-old systems. The patch, in other words, did the damage.

The pressure has an identifiable source. Anthropic released Claude Mythos Preview in April, a frontier model promising autonomous discovery of thousands of zero-days, meaning flaws unknown to the vendor and therefore unpatched. Mozilla's Firefox developers patched fewer than 30 bugs a month in 2025 but shipped over 400 fixes in April this year, most of them found by Mythos Preview. Microsoft's Patch Tuesday, its monthly security update, went from fewer than 100 fixes before April to patches for more than 1,000 vulnerabilities across the past two months.

That volume is what teams now have to absorb. The steps below show how to do it without repeating the outage week.

What you need before starting

  • Difficulty: intermediate. Time to set up: roughly a week for a small estate, longer for a large one; thereafter a recurring monthly cycle.
  • An asset inventory that shows which systems are internet-facing, which run legacy infrastructure, and which have critical dependencies.
  • A patch management or endpoint tool that can target groups of devices rather than only the whole estate.
  • A test or pilot group that mirrors production, including at least one legacy system.
  • A rollback method for every system class, whether snapshots, backups or vendor uninstall procedures.
  • A named owner for the change-management decision, so the call to patch fast or hold does not default to whoever is loudest.

Step-by-step process

Step 1: Rank every new patch by actual risk (about 30 minutes per cycle)

Dan Schiappa, president of technology and services at Arctic Wolf, puts the tension plainly: vulnerabilities being actively exploited often demand rapid remediation, because a security incident can have far greater consequences than an IT disruption. Start there. Sort incoming fixes into three buckets: actively exploited, internet-facing but not known to be exploited, and everything else.

The data supports this triage. Patrick Garrity, a security researcher at VulnCheck, has pooled Anthropic-discovered vulnerabilities from a range of sources. In the seven months since Project Glasswing launched, he counts 442 newly assigned CVEs discovered using Claude models. Only 0.45% are being exploited in the wild, namely CVE-2026-61500 and CVE-2026-26980. A thousand-patch month is not necessarily a thousand emergencies. Pieter Arntz, senior intelligence reporter at Malwarebytes, says teams should prioritise according to the actual risk and impact of exploitation, and that limiting the options to leaving everything vulnerable or patching everything blindly is wrong.

Step 2: Map the blast radius of each fix (one to two hours)

For anything in the top bucket, list what depends on the system being patched. Schiappa warns that in complex environments, particularly those running legacy infrastructure, rushed patching can overload systems, create compatibility issues, disrupt critical dependencies and trigger outages that affect business operations. A dependency map turns that warning into a checklist: which services call this one, which batch jobs touch it, and what breaks if it restarts.

Step 3: Define your own line between speed and uptime (half a day, once)

Shane Fry, CTO of RunSafe Security, notes that the line between preemption and self-destruction differs for each company. Some prioritise data security; others absolutely must avoid downtime. Write down, per system class, which side you lean towards and who can override it. Fry's framing is that security should not be treated as a choice between immediate patching and business continuity. The goal is reducing risk without creating new operational failures.

Step 4: Deploy in phases (days, not hours)

Schiappa's guidance is that patches should be rolled out in a phased manner so teams can monitor behaviour, validate effectiveness and identify unintended impacts before broader deployment. A workable sequence:

  1. Pilot group, including a legacy system.
  2. A small slice of production.
  3. The remaining estate.

For actively exploited flaws, compress the timeline; for the rest, let it breathe. Just because you can issue a patch immediately, Schiappa says, does not mean it should be deployed across the entire environment at once.

Step 5: Monitor, validate and record (ongoing)

After each phase, check service health, error rates and dependent systems before moving on. Log what was patched, when, and any side effects. That record is what lets you say no to the next panic with evidence.

Common mistakes to avoid

  • Treating discovery volume as exploitation volume. The Garrity figures show most AI-found bugs are not being used by attackers, at least not yet.
  • Pushing one patch to the whole estate on day one. This is the pattern behind the outage week.
  • Skipping legacy systems in the pilot. Decades-old systems are where compatibility problems surface.
  • Applying no rollback plan. A patch you cannot undo turns a bad day into an incident.
  • Letting a vendor's urgency set your timeline. Vendors are themselves overwhelmed; many open-source projects have shut down or paused bug bounty programmes because they could not keep up.

Tips that make a difference

  • Watch disclosure timelines. Google's Project Zero is cutting its disclosure timeline from 120 days to 90, and to 30 days by 2028, fearing researchers or attackers will use flaws before patches ship. Shorter windows mean less warning, so rehearse your fast path now. The site's guide to getting patch-ready for shorter disclosure windows covers that preparation.
  • Compensating controls buy time. Where a patch cannot go in immediately, isolation or access restrictions on the affected system reduce risk while testing completes. This is a practical inference from the risk-reduction goal, not a vendor prescription.

UK angle

For UK organisations, the same discipline maps onto existing obligations. A botched patch that takes systems down is an availability failure as well as an operational one.

Next steps

  • Run a tabletop exercise on a failed emergency patch.
  • Review your backup and rollback readiness; the site's UK guide to business backup and recovery is a good starting point.
  • Track exploitation data monthly and adjust your buckets accordingly.
  • Revisit the process when Project Zero's disclosure window shortens.

A growing consensus holds that the outage week was caused by how organisations responded to a flood of AI-found vulnerabilities, not by the flood itself. Jason Healey, a cyber risk academic at Columbia University and former White House cyber policy director, says there is no cheat guide because this is the first time anyone has done it. The practical answer is a process, not a reflex.

A risk-ranked, phased patching routine can be summarised simply: triage by exploitation and exposure, map dependencies, set your own speed-versus-uptime line, deploy in stages, and monitor each one. Adopt it before the next thousand-fix Patch Tuesday arrives.

Frequently asked questions

Should organisations patch everything immediately?

No. Actively exploited flaws need rapid remediation, but patches should be rolled out in phases so teams can monitor behaviour, validate effectiveness and spot unintended impacts before broader deployment.

What caused the recent week of IT outages?

A growing consensus holds that fear of AI-powered threats, not AI attacks themselves, caused them, through forced upgrades and patches applied to decades-old systems.

How many AI-found vulnerabilities are actually exploited?

VulnCheck's Patrick Garrity counted 442 newly assigned CVEs discovered using Claude models in seven months, of which only 0.45% are being exploited in the wild.

How should teams prioritise patches?

By actual risk and impact of exploitation, according to Malwarebytes' Pieter Arntz, rather than leaving everything vulnerable or patching everything blindly.

More from Future Technology HQ