Start your day with intelligence. Get The OODA Daily Pulse.

Home > Analysis > OODA Original > Rigging for Stormy Weather: Preparing for Outages, Constant Patching, and AI-Enabled Attacks

Rigging for Stormy Weather: Preparing for Outages, Constant Patching, and AI-Enabled Attacks

Mariners do not wait for the storm to prepare for heavy seas. Much of the technology economy has spent two decades doing the opposite, stripping out slack in the name of efficiency, and the result performs beautifully in fair weather and fails in surprising ways when the sea changes. This post sets out the situation, grounded in evidence, and then offers advice for individuals and companies. It engages two extremely prescient OODA Loop pieces from Patrick Scannell that frame the problem, The Architecture of Normal Collapse and Guarding the Wrong Door, and ties them to the operational data on outages, patching, and AI-enabled attacks.

Bottom Line Up Front

  • Three currents are converging. Overoptimized, tightly coupled systems produce outages with no attacker involved. The patch treadmill is accelerating due to AI enabled tools finding flaws in our systems, and this itself is a source of both risk and outages. And AI is compressing the time attackers need to find, weaponize, and chain weaknesses.
  • The house is the hazard, and the monsters read the house. As The Architecture of Normal Collapse argues, intruders and AI do not create fragility. They read it, and some of them pile on.
  • The answer is preparation, not panic. Build resiliency on purpose, measure how long you can run on your own, and rehearse the combination of failures rather than one failure at a time.

The Situation: Three Currents, One Sea

Current one: overoptimization and the loss of resiliency

The central claim of The Architecture of Normal Collapse is that a habitat built from accelerating technology and dense interdependence may be outrunning the ability to repair it. Drawing on Charles Perrow’s work on normal accidents, the paper argues that when interactive complexity and coupling grow faster than slack and diagnostic capacity, ordinary shocks cost more and last longer. Among its indicators is cloud concentration: the three largest providers took 63 percent of enterprise cloud spending in Q3 2025, up from 61 percent two years earlier.

Its companion, Guarding the Wrong Door, makes the attention argument. Using Pearl Harbor, where aircraft were parked wingtip to wingtip to guard against sabotage, it shows how a vivid threat can crowd out a larger one coming through an unwatched door. The paper grants that AI is a real monster under the bed, and argues that the house it stands in is the bigger problem. It also warns against overcorrecting, so the advice below is deliberately proportionate.

The operational record backs the thesis. Uptime Institute’s 2026 annual outage analysis finds that per-site outage rates have declined for a fifth straight year, which is good news, but also that the improvement is slowing, that external infrastructure failures are becoming more prominent, and that resiliency assessments remain more focused on internal systems than on external and systemic risks. Uptime expects failures to be increasingly linked to complex interactions between systems, and notes that more automation can cause different classes of problems.

Three incidents show what that looks like in practice, and none of them involved an attacker:

  • CrowdStrike, July 2024. A content update to a security product crashed Windows machines worldwide, affecting an estimated 8.5 million devices according to Microsoft, as relayed by CISA. CrowdStrike’s own root cause analysis traced the crash to a mismatch between 21 input fields defined for a new template type and 20 values supplied by the sensor code, which evaded validation and testing. OODA Loop’s analysis of the lessons on harmonization and strategic risk traces the aftermath.
  • AWS us-east-1, October 2025. In its post-event summary, AWS described a latent race condition in the automated DNS management system for DynamoDB that left an incorrect empty DNS record the automation failed to repair. The event ran from 11:48 PM PDT on October 19 to 2:20 PM PDT on October 20, and dependent services including EC2 launches and Network Load Balancer were affected.
  • Cloudflare, November and December 2025. Cloudflare’s post-mortem says a database permissions change doubled the size of a Bot Management feature file past a software limit, and that no cyberattack was involved. Its Fail Small plan notes a second failure on December 5 that affected 28 percent of applications for about 25 minutes. OODA Loop’s brief at the time noted the breadth of the disruption.
  • November 2026, the outages reach mass media. As we capture in In an Over-Optimized World, Normal Accidents Are Inevitable, in a single week, Microsoft 365, JPMorgan Chase, the FAA and the NSA all suffered outages. Senate Intelligence Committee Vice Chair Mark Warner said current assessments show no sign of foreign interference or coordinated attack, and blamed the unintended consequences of rapid software patching instead. Experts including former Deputy National Security Adviser Anne Neuberger warned that rushing out patches for flaws found by AI overwhelms legacy systems never built for that pace, causing cascading failures. As OODA Loop reported, citing the Financial Times, this was the first time the problem broke into mainstream news.

A fourth case involves AI. The Financial Times reported that Amazon’s Kiro coding tool deleted and recreated an environment, causing a 13-hour disruption in December 2025, as OODA Loop summarized. Amazon told the FT, as reported by Tom’s Hardware, that this was user error, not AI error. Either way, autonomous tools with production permissions and no second approval are a new kind of coupling.

Current two: the patch treadmill

Patching is the right answer to known vulnerabilities, and it is also becoming a continuous operational burden. Verizon’s 2026 Data Breach Investigations Report found that exploitation of vulnerabilities is now the most common initial access vector at 31 percent of breaches, up from 20 percent, and ahead of credential abuse. In the same dataset only 26 percent of vulnerabilities in the CISA Known Exploited Vulnerabilities (KEV) catalog were fully remediated in 2025, down from 38 percent, and the median time to full resolution rose from 32 to 43 days. Verizon’s announcement adds that threat actors are using AI to shorten the time to exploit known vulnerabilities.

Mandiant’s M-Trends 2026 estimates the mean time to exploit at negative seven days, meaning exploitation is routinely occurring before a patch is even released. That figure is an average drawn from Mandiant’s incident investigations and should be read as direction rather than a countdown for any one flaw, but the direction is clear. Meanwhile the supply of newly found flaws is growing. OODA Loop reported that Anthropic’s Mythos Preview effort had uncovered more than 10,000 high or critical vulnerabilities, with Anthropic saying that progress is now limited by how quickly findings can be verified, disclosed, and patched. Anthropic’s disclosure dashboard reports, as of November 2, 5,087 vulnerabilities disclosed across 680 open source projects, of which 629 were known to be patched, and says human triage is the rate-limiting step. Anthropic is also shortening its disclosure windows to 30 days for high and critical issues from January 1, 2027, while VulnCheck’s November analysis argues the program is overwhelming maintainers with findings of uneven value. On either view, more discovery and shorter windows mean more patches arriving faster, for defenders who already struggle to keep up.

There is a second edge to this sword. The CrowdStrike and Cloudflare outages were both caused by updates pushed to fix or improve things. Patching at speed without staged rollout turns the cure into an outage. And as David Bray and Jeff Frazier argue in Why Patching Vulnerabilities Is No Longer Enough, patching alone is incomplete in any case: it cannot cover systems nobody has scanned, and it does nothing about attackers who abuse legitimate tools. They make the case for detecting malicious intent in real time through patterns-of-life analysis.

Current three: AI-enabled attackers

Google Threat Intelligence Group’s November 16 AI Threat Tracker says adversarial AI has reached an inflection point. It reports threat actors abusing open source agentic penetration testing frameworks, naming ARTEX, CAIRN, CyberStrikeAI, and PentAGI, in live intrusions, and using LLMs to automate patch reverse engineering through differential analysis so that n-day vulnerabilities can be weaponized quickly and exploitation scaled against internet-facing servers without human intervention. The FBI’s FLASH on ARTEX, analyzed in FBI FLASH on ARTEX, reached a similar conclusion from the law enforcement side.

Claude Mythos Shows Why Technological Advancement Outpaces Risk Mitigation argues that autonomous vulnerability research favors the aggressor, and China’s Mythos Moment reports that Chinese firm 360 Security says it has built a comparable system, though its vulnerability counts have not been independently verified.

AI-enabled attackers were not needed for the over-optimization related outages in November, but attacks sure are up. A few examples: a $25 million deepfake video-call fraud at Arup, North Korean IT-worker schemes scaled with generative AI and Anthropic’s disruption of a China-linked espionage campaign run largely by Claude Code. It also includes a ransom intrusion compressed into under 10 hours by AI agents, Storm-0940’s ARTEX-driven ConnectWise supply chain breach, the agentic attack that exposed every Chainalysis user, a $129 million AI voice and image impersonation scheme charged by federal prosecutors and an AI-tool-assisted attack on the European Commission. The two fraud cases alone account for more than $150 million in direct losses. With response, downtime and breach costs at the other victims still undisclosed, the true total is likely far higher.

Why the combination matters

Each current is manageable alone. The danger, as The Architecture of Normal Collapse puts it, is the stacking. A fragile house changes the adversary’s problem: they do not have to build the weakness, only find the trigger nobody is watching, and AI lowers the cost of finding it. A vendor outage, an emergency patch cycle, and a wave of machine-speed exploitation can land in the same week. That is why the advice below leans on slack, small failure domains, and rehearsal of compound events, the measures Guarding the Wrong Door recommends: keep slack on purpose, run controlled burns, and draw the boundary around the whole system and not just the door.

None of this is a forecast of collapse. Uptime’s data show digital infrastructure remains remarkably resilient, and the OODA papers present their hypothesis as testable. But preparation is cheap relative to the cost of being wrong, and most of it pays off in fair weather too.

Advice for Individuals

The Architecture of Normal Collapse offers a useful ladder, the rule of threes, which extends the familiar survival rule to what systems can deliver: three days for water, light, and communication, and three weeks for food, medicine refills, and local services. Households can cover the first rungs and help their communities cover the next. Start there.

Cover the first 72 hours and then some

  • Water, food, and Batteries. Ready.gov advises storing at least one gallon of water per person per day for several days. Build a basic emergency kit that includes a phone charger and backup battery, and extend it as your situation requires.
  • Medicine. Ready.gov notes that about half of Americans take a prescription medicine daily and that an emergency can make refills difficult. Ask the medication-clock question: how many days of each daily medicine do you have on hand? Raise the lowest number.
  • Cash and paper. Ready.gov’s kit list includes cash and copies of important documents. Payment systems, ATMs, and online banking are among the first things users notice when a cloud or network provider has a bad day, and Ready.gov notes that power outages can close ATMs, banks, gas stations, and stores.
  • A plan that does not depend on the network. Write down key phone numbers and meeting points, and agree who calls whom. Use the Ready.gov planning guidance as a template, and rehearse a weekend without power and without the router.

Make your digital life update-proof and attack-resistant

  • Turn on automatic updates, and keep a recovery path. CISA’s Secure Our World guidance recommends automatic updates because many fix security risks. Pair that with backups of the things you cannot replace, kept somewhere an update or an attacker cannot reach, such as an external drive stored offline. For organizations CISA recommends offline, encrypted backups that are regularly tested, and the same logic applies to a household’s photos and records. Test a restore once.
  • Use multifactor authentication on the accounts that matter. CISA explains that MFA protects an account even if a password is stolen. Prioritize email, banking, and your password manager.
  • Replace unsupported home network gear. Verizon’s DBIR reports that VPNs, firewalls, routers, and switches accounted for 5 percent of breaches in its dataset, up from 1.5 percent a year earlier, though Verizon notes part of that change reflects how it codes remote access cases. Treat an old router as a liability and enable its automatic updates.
  • Assume convincing fakes. AI lowers the cost of tailored phishing and impersonation, and Verizon names pretexting as its second most frequent social action. Agree a code word with family and a callback rule at work: any urgent request for money or credentials is verified through a second channel you chose, not the one the request arrived on.

Build skills and relationships

  • Learn one skill that works without power and teach someone else. Basic first aid, simple repair, and running your home without the grid are good starting points.
  • Know your neighbors, and keep proportion. The first help in most disasters comes from the people nearest you. As Guarding the Wrong Door reminds us, fear can drive choices that hurt more than the threat itself, so a few inexpensive steps beat anxiety.

Advice for Companies

Executives hold the rungs that households cannot: the three-month and three-year rungs, where supply, replacement, and recovery capacity live. The OODA paper’s own question for company leaders is blunt: how many days can we run with no outside help? The measures below turn that question into action. They align with NIST Cybersecurity Framework 2.0, CISA’s Cross-Sector Cybersecurity Performance Goals, and NIST’s contingency planning guidance.

Every company is unique of course. Consider the below just food for thought and build out your own plan by internal focus groups and look for gaps in your planning by use of tabletop exercises.

Measure your days of autonomy and map your dependencies

  • List the ten or so services that must not stop, and for each map what sits underneath: cloud region, DNS, identity provider, endpoint security agent, payment processor, key SaaS tools, carriers, and the people who can run the process by hand.
  • Set an impact tolerance for each service and test it. Uptime found that resiliency assessments focus on internal systems more than external dependencies, which is the gap to close. If one provider, region, or vendor update channel can stop a critical service, that is a single point of failure whatever the contract says.
  • Report a few numbers to the board: days of autonomy per service and the date last tested, share of services with tested failover, median time to remediate known exploited vulnerabilities, tested time to restore the top ten systems, and the share of changes and vendor updates deployed through staged rollout.

Buy back slack on purpose: Invest in resiliency

  • For crown-jewel services, fund real redundancy: a second region or provider that is tested, not just contracted. It will look inefficient on a quarterly report. That is the point.
  • Buy redundant comms from different providers
  • Consider where you will want starlink.
  • Keep a manual mode and the people who know it, with a written degraded mode for each critical service and clear decision rights on who can switch it on. Pool spares of slow-to-replace items with peers.

Treat your own change process as a threat

  • Stage everything. CrowdStrike’s root cause analysis lists staged deployment with canary testing and customer control over content updates among its mitigations. Cloudflare’s Fail Small plan commits to health-mediated deployments for configuration changes, in which each service defines the success metrics, rollout plan, and rollback steps. Apply the same discipline to your own patches and changes, and ask vendors that push updates into your environment for staged rollout and customer control.
  • Limit blast radius. Aim for failure domains small enough that a bad change takes out a slice, not the whole estate. AWS’s post-event summary describes a new velocity control limiting how much capacity a single load balancer can remove during failover, the same idea.
  • Put guardrails on AI agents with production access. Give agents least-privilege credentials, require a second human approval for destructive or production-wide actions, and log what they do. OODA Loop’s playbook for corporate leaders on agentic AI risk and the OODA Network survey on agentic AI cover the risk classes in more depth.

Run a risk-tiered patching lane, with an emergency path

  • Recommendation: Start with known exploited vulnerabilities. Treat the CISA KEV catalog as your first filter. Its parent directive, BOD 22-01, requires federal civilian agencies to remediate within two weeks for newer entries, and it is a reasonable benchmark even if you are not bound by it. Against Verizon’s 43-day median, that gap is the number to close.
  • Prioritize internet-facing and edge systems. Mandiant reports espionage groups deliberately targeting VPNs and routers that lack endpoint detection telemetry. Inventory them, patch them first, and retire end-of-life gear.
  • Pre-authorize an emergency lane. Decide in advance who can approve an out-of-cycle patch, how it is staged, and what you do when there is no patch: isolate, disable the feature, add compensating controls, and increase monitoring. Google’s report describes attackers automating patch reverse engineering, so your approval cycle should not take weeks.
  • Do not rely on patching alone. Invest in behavioral detection that spots abuse of legitimate tools, the Bray and Frazier point, since those attacks need no unpatched flaw.

Defend at machine speed, without adding fragility

  • Shorten detection and response times and pre-approve automated containment for high-confidence cases, such as isolating a host or disabling a compromised account, with a human in the loop for the rest. NIST’s incident response guidance is a good structure for playbooks.
  • Harden identity first: phishing-resistant authentication for privileged and remote access, limited standing privileges, and segmentation so one foothold does not become a campaign. Test your controls against the mitigations in our ARTEX analysis.
  • Judge every new safeguard, defender-side AI included, on whether it adds slack or adds complexity. That is Perrow’s caution as repeated in the OODA paper, and it is a good procurement question.

Make recovery real, and rehearse the combination

  • Maintain offline, encrypted, immutable backups and test restores on a schedule, as CISA’s #StopRansomware guide recommends, with golden images for fast rebuilds. Measure how long it takes to rebuild your most important services and publish that number internally. Keep out-of-band communications for the crisis team, and test them.
  • Run compound exercises: a vendor update takes down endpoints, a high-severity exploit is announced the same day, and an impersonated executive messages finance with urgent instructions. The dangerous case is a material failure arriving in a contested information environment.
  • Run small controlled burns, such as failing over a service or revoking a credential on a schedule, and publish what failed. Include critical suppliers: ask for their recovery times and update practices, and write notification and staged-update expectations into contracts.

What This Means for Executives

  • Efficiency is a decision, not a default. Every layer of slack you remove is a bet that nothing will go wrong at the same time. Make that bet explicitly.
  • Your change process is part of your attack surface. The largest recent outages came from routine updates, not adversaries. Staged rollout and blast radius limits are as important as any security tool.
  • Patching is necessary and no longer sufficient. Run it on a risk-tiered clock with an emergency lane, and pair it with detection, identity hardening, and tested recovery.
  • Prepare for the combination, and keep proportion. Exercise vendor failure, active exploitation, and a contested information environment together. The goal, in the spirit of the OODA papers, is not to brake progress but to steer it toward a more durable kind. Carry the ballast, and keep moving.

Further Reading on OODA Loop

Bob Gourley

About the Author

Bob Gourley

Bob Gourley is an experienced Chief Technology Officer (CTO), Board Qualified Technical Executive (QTE), author and entrepreneur with extensive past performance in enterprise IT, corporate cybersecurity and data analytics. CTO of OODA LLC, a unique team of international experts which provide board advisory and cybersecurity consulting services. OODA publishes OODALoop.com. Bob has been an advisor to dozens of successful high tech startups and has conducted enterprise cybersecurity assessments for businesses in multiple sectors of the economy. He was a career Naval Intelligence Officer and is the former CTO of the Defense Intelligence Agency.