# I Was Told My Cyber Risk Keynote Was Too Alarmist. Then October 21 Happened.
By 

## Article Content
Six cybersecurity stories. One extraordinary 24-hour period. And a rather uncomfortable validation of something I said on stage in August. Yesterday, every alarm on the internet seemed to go off at once. In just 24 hours: 7 in 10 websites run on PHP, the software underneath WordPress, and the US National Security Agency warned organizations to take their systems offline if they couldn't fix two critical flaws right away. Thousands to millions: Cloudflare saw thousands of attempts to exploit those PHP flaws before they were even announced. Within hours, it was millions. 10 out of 10: GitLab, where companies build and store their software, disclosed a flaw with the highest possible severity score, hiding in 12 years of releases. 2,250+ cloud environments were targeted by a China-state-aligned group using AI agents to automate its attack, according to Microsoft. 2,900 login attempts a minute. That's how fast the AI-driven attack moved. Under 3 hours: in some cases, that's all it took to go from first contact to the keys to an organization's cloud. And more: emergency fixes for Node.js, another building block of the web; a warning from cloud platform Railway that customers' passwords and digital keys may have been exposed; and new research showing how malicious code can hide inside the training of AI. A few people told me my August keynote was too alarmist. Yesterday, reality ran the script — at machine speed. In August, I delivered the closing keynote at the 2026 Governance, Risk and Control Conference in San Diego, jointly presented by ISACA and The Institute of Internal Auditors. My message was built around a concept I've been developing for some time: the Acceleration Gap. The idea is straightforward. The speed at which risk emerges, evolves, spreads and compounds is accelerating faster than the governance structures organizations have developed to understand and manage it. We are entering a world in which threats can move at machine speed while our ability to assess, audit, control and respond to them remains stubbornly stuck in human-speed processes. On August 19, the morning of my keynote, I published an article outlining precisely what I was about to tell the audience. I wrote: "Cyber threats now move in minutes — sometimes seconds." My article included numbers that I found rather startling: a 263% increase in reported vulnerabilities over five years, attack breakout times measured in minutes and even seconds, and an extraordinary gap between the time it takes attackers to exploit a vulnerability and the time organizations require to patch it. It was a rather blunt keynote. Deliberately so. The overwhelming majority of the audience loved it. My speaker rating was 4.69 out of 5, with more than 92% positive feedback. But a few people thought I'd gone too far. The picture I painted was too dark. The message was too alarming. Perhaps, in their view, I was being a little too shrill about the future of risk. I took those comments seriously. I always do. Then, in September, Anthropic published its threat intelligence report, documenting how artificial intelligence was already being used to orchestrate increasingly sophisticated attacks. I wrote about that uncomfortable validation of my keynote on September 11, in an article titled I Was Told My AI Risk Keynote Was Too Dark. Then Anthropic Published the Footnotes. My conclusion? "My reaction was not that I had overstated the risk. It was that I might have understated the speed." And now we have October 21, 2026. What a day. When six different parts of the digital world start flashing red Consider what happened in one remarkable news cycle. First, the US National Security Agency issued an extraordinary warning about PHP, the programming language used behind the scenes by an enormous proportion of the world's websites. Two critical vulnerabilities could allow attackers to execute code on vulnerable servers without authentication. The warning was sufficiently serious that organizations unable to patch immediately were advised to consider taking affected systems offline. European cybersecurity agencies issued coordinated warnings, and websites were taken offline as a precaution. And get this: AI-assisted research helped discover these vulnerabilities. The same technology transforming our ability to uncover security flaws is also increasing concern about how quickly such weaknesses can be weaponized. Think about that for a moment. One of the fundamental building blocks of the modern Web was suddenly the subject of urgent international security alerts. This wasn't some obscure piece of software running in a forgotten corner of a data centre. This was part of the plumbing upon which much of the digital economy depends. How quickly did the attackers move? Late on October 21, Cloudflare co-founder and CEO Matthew Prince shared what his company was seeing across its network: "Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now." Thousands of exploit attempts before the vulnerability was even publicly announced. Millions within hours of the announcement. That's exactly what I showed the audience in August: the time from patch to exploit has collapsed from 63 days, to 5 days, to negative. Sometimes the attackers are already running before the defenders hear the starting gun. Then came GitLab. A newly disclosed vulnerability, CVE-2026-110264, received a maximum severity score of 10.0. The flaw, which could allow unauthenticated remote code execution under certain conditions, affected self-managed GitLab releases stretching all the way back to 2014. Twelve years of software history suddenly became part of an urgent patching exercise. Then Node.js. A critical component of modern web application infrastructure faced its own emergency security updates, including a high-severity flaw in its Brotli decoder that could potentially permit remote code execution. Vercel announced that its web application firewall rules were blocking known exploit patterns, while also making clear that those protections were no substitute for patching vulnerable software. PHP. GitLab. Node.js. Three different foundations of modern computing. Three major security events. One day. And we were just getting started. The supply chain becomes the attack surface Railway, the cloud application deployment platform, disclosed a security incident involving systems used to store customer environment variables. Its preliminary investigation indicated that an unauthorized party might have accessed encrypted variables and, in some cases, obtained their decrypted contents. That sounds rather technical until you understand what those variables can contain. API keys. Database credentials. Registry access tokens. The digital keys that allow applications to communicate with other applications, reach customer information and operate across multiple interconnected services. Railway recommended that users rotate their credentials as a precaution while its investigation continued. Here was an almost perfect demonstration of something I emphasized in August: ecosystem risk. In my September article, I described how risks increasingly extend beyond the walls of any individual organization. They can emerge from a vendor, a software update, an exposed credential, a third-party connection or an API that suddenly becomes a gateway into hundreds of other systems. Your organization might do everything right with its own infrastructure. Your internal controls might be excellent. Your cybersecurity policies might be exemplary. But your systems are increasingly dependent on platforms, libraries, services, credentials and software components that you neither own nor fully control. Your risk is no longer limited to your technology. It's embedded in everyone else's technology too. And now even the mathematics of AI becomes a potential attack vector On October 20, security researcher Elon Litman published a fascinating piece of research involving PyTorch, one of the fundamental software frameworks used to build and train artificial intelligence systems. His demonstration showed how apparently harmless code added to an AI training process could exploit differences between floating-point arithmetic and automatic differentiation, causing computations to execute through the model's training process while reporting a zero-valued contribution to its loss function. That's a mouthful. Here's the simpler version: something that appears to do absolutely nothing could potentially be designed to do something remarkably consequential. This was a security research demonstration, not evidence of a widespread attack. But it exposed another category of supply-chain risk that organizations will increasingly need to understand. Think about the implications for AI governance. We are building enormously powerful systems using increasingly complex combinations of models, code, mathematical methods, open-source packages and third-party tools. How many organizations can genuinely tell you that they understand every component in that chain? And how many have governance structures capable of keeping up with the new risks being discovered inside it? Then Microsoft supplied the most extraordinary evidence of all On October 21, Microsoft Threat Intelligence published research about a China-state-aligned threat actor it tracks as Storm-0940. According to Microsoft, the group had been using an agentic AI penetration testing framework called ARTEX to automate attacks targeting hundreds of organizations around the world. Microsoft linked the activity to the recently disclosed ConnectWise supply-chain compromise. The numbers are astonishing. Microsoft documented attack activity reaching more than 2,250 customer environments, with authentication attempts occurring at rates of approximately 2,900 per minute. It reported multiple cases in the broader campaign in which the time between initial targeting and privileged access to cloud infrastructure was less than three hours. The majority of targeting attempts were unsuccessful, an important distinction. But in a number of cases, attackers achieved privileged access and established opportunities for persistent control. Now compare those numbers with the governance processes in the average large organization. How long does it take to schedule a risk committee meeting? Approve a policy revision? Complete a third-party security review? Escalate an incident through the established chain of command? The attackers are operating in minutes. Sometimes the defenders are still operating in months. That is the Acceleration Gap. And it's precisely what I was talking about in August. In my original August 19 article, I compared the extraordinary speed of modern attacks with the much slower processes organizations use to assess and manage them. I argued that the lifecycle of a threat was becoming shorter than the lifecycle of the governance process designed to address it. That was the fundamental message. October 21 gave us another extraordinary illustration of just how significant that disconnect has become. The Internet is a giant game of Jenga. And someone keeps removing blocks. There's a wonderfully chaotic cartoon making the rounds showing the entire digital infrastructure of the modern world balanced precariously on a collection of open-source projects, programming languages, cloud platforms, WordPress, Linux, PHP, DNS and a handful of exhausted developers. A remix of Randall Munroe's xkcd comic "Dependency" (xkcd.com/2347) It's funny because it captures something profoundly important. We've created an extraordinary digital civilization through layer upon layer of technology, much of it developed independently, maintained by small teams, connected through elaborate dependencies and trusted by organizations that have only limited visibility into what lies underneath. It's an engineering miracle. It's also an extraordinarily complex structure of interdependent risk. And now we're adding AI agents to the top of the pile, giving increasingly autonomous software the ability to access systems, make decisions, initiate transactions and execute actions at machine speed. Some of these developments are spectacular opportunities. Others create entirely new vulnerabilities. Often, the very same technology does both. That's why the cyber events of October 21 matter so much. They weren't all the same type of event. Some were newly disclosed vulnerabilities, one was an active security incident, one was a research demonstration and another was evidence of an actual AI-accelerated attack campaign. But taken together, they illustrate how many different directions risk can emerge from, how deeply interconnected our technology has become, and how little time organizations may have to react. We can't keep governing tomorrow's risk with yesterday's assumptions One of the most important slides in my August keynote carried a rather pointed observation: “Risk now moves faster than governance can certify it.” I wasn't suggesting that the risk, audit and compliance professions had failed. Far from it. I was suggesting that the fundamental assumptions underlying their work were being challenged by an entirely new reality. We built governance systems around predictable reviews, documented processes, periodic assessments and established control frameworks. Those processes remain important. But too often, they assume that organizations have sufficient time to identify a problem, understand it, analyze the implications and then develop an appropriate response. Increasingly, they don't. When a single software vulnerability can generate worldwide security warnings, when compromised credentials can expose interconnected cloud environments, and when AI agents can automate thousands of hostile actions in minutes, the old governance timetable becomes part of the problem. That's not an argument for abandoning governance. It's an argument for reinventing governance. More continuous assurance. More real-time visibility. Better understanding of digital dependencies. Faster decision-making. Greater emphasis on resilience, not just compliance. And, perhaps most importantly, a willingness to explore risks that haven't yet made it onto the official risk register. We need to stop asking only whether our organizations are compliant with the risks we already understand. We need to start asking whether we're learning quickly enough to recognize the risks we don't. So, was my keynote too alarmist? I keep coming back to those few comments from August. I understand the reaction. Nobody particularly enjoys being told that the systems, controls, expertise and assumptions upon which they've built a career might be facing an extraordinary period of disruption. But there's a difference between being alarmist and sounding an alarm. An alarmist exaggerates danger beyond the available evidence. A futurist has a different responsibility: to identify emerging patterns, connect developments that might initially seem unrelated, challenge assumptions and help organizations prepare for what comes next. The point isn't to make people afraid of the future. It's to help them avoid being surprised by it. And that's why I find the events of October 21 so consequential. They don't prove that every scenario I outlined in August will happen, or that every organization faces every one of these threats. But they provide remarkable evidence of the underlying direction of change. Risk is accelerating. Our digital dependencies are multiplying. AI is changing the economics of attack. The time available to understand and respond is shrinking. All of that was central to my August keynote. I ended that presentation with a challenge built around four capabilities: exploring, learning, experimenting and anticipating. Organizations need to become structurally better at all four because the half-life of what we know about risk is shrinking. That is still the message. In 2016, the conversation was about auditing at the speed of risk. In 2026, it is about governance at the speed of risk. And looking toward 2036? Governance at the speed of machines. That was the closing idea of my August presentation. In my September follow-up, I wondered whether I had painted too dark a picture or whether the future was simply moving faster than the audience expected. Now, on October 22, we have another remarkable set of developments that make the question even more relevant. Some people might still think that message is too alarming. But consider the chronology. August: I warned that the speed of risk was becoming unmanageable within traditional governance structures. September: Anthropic documented how AI was changing the scope, complexity and economics of cyberattacks. October: A cascade of major cybersecurity disclosures, alongside Microsoft's documentation of real-world, AI-accelerated attack activity operating at machine speed. I don't particularly enjoy writing articles about how quickly emerging threats are becoming operational realities. I'd much rather be writing about the extraordinary opportunities that come with the same technologies. And I do, every single day. But the future doesn't limit itself to the opportunities we find comfortable. The risk isn't that we're sounding the alarm too loudly. It's that the future is arriving faster than many organizations can hear it. Earlier articles by Jim Carroll: August 19, 2026: The Acceleration Gap: Why Risk Now Moves Faster Than Governance Can Keep Up September 11, 2026: I Was Told My AI Risk Keynote Was Too Dark. Then Anthropic Published the Footnotes. Technical sources: NSA PHP advisory · GitLab security release · Node.js security release · Railway disclosure · Elon Litman's PyTorch research · Microsoft Threat Intelligence · PHP2Shell disclosure · Vercel Node.js bulletin · Matthew Prince on X.

---
Source: https://jimcarroll.com/2026/10/i-was-told-my-cyber-risk-keynote-was-too-alarmist-then-october-21-happened/