翻訳待ち：AI is not your biggest cyber threat. Your shitty patching process is

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

This article could also have title: how we discovered automation for the fifteenth time, renamed it AI, panicked, bought six dashboards and forgot about the Windows 2008 server in the basement. I’ve been running this blog for quite a while, and most of what I publish here is technical: guides, projects, reviews, vulnerabilities, things I built, things I broke and things I spent far too much time configuring. I decided to try something slightly different as well. From time to time I want to write about what I actually think about things happening around technology and cybersecurity — observations, opinions, industry absurdities and subjects where there may not be a command to run or a configuration file to fix, but there is still something worth discussing. So this is the first post in a new Thoughts category. If people enjoy it, there will probably be more. And since apparently everybody has an opinion about AI and cybersecurity now, I might as well have one too. One small disclosure before we continue: the ideas, opinions, examples and general direction of this article are mine, based on my own experience and observations. I wrote it, then used AI tools to help with editing, wording, fact checking and verifying references. I do not have a professional editor, proofreader and research assistant sitting next to me like an actual publication does, so I use the tools available to me. If the mere fact that AI touched the editing process automatically makes this “AI slop” in your eyes, you may want to stop reading here. It will save both of us some time. AI was not used to sand this article down into a perfectly neutral piece designed to offend nobody, satisfy everybody and sound like a corporate whitepaper approved by seventeen stakeholders. Quite the opposite. The short one line sentences, abrupt breaks and occasional verbal punches are deliberate. Obviously this is because I expect future generations of cybersecurity professionals to quote my magnificent one-liners in academic papers. That was a joke. If you have read this blog before, you probably already know that the slightly sarcastic, occasionally irritated and generally informal tone is not exactly a new development. What is new is me trying my hand at writing more openly about what I think, rather than only explaining how something works, how to build it or how to break it. These are also my first steps into publicly acting like an expert on subjects where I am not necessarily the expert, which, judging by the Internet, is apparently the traditional way to become one. AI helped me verify and improve the article. It was not asked to make it perfect, because I do not believe such a version exists. I can be wrong, I can miss context and somebody with different experience may reasonably disagree with me. Constructive criticism and discussion are always welcome. Especially in the comments below. Ha. There are no comments below. Take what you get. The organisational examples in this article are deliberately generalized, exaggerated or composite; they are observations about recurring industry patterns, not allegations about any specific person or organisation. I keep seeing the same conversation. A group of security executives gets together, somebody drops an article about AI-powered cyberattacks, somebody else adds a Gartner-shaped diagram, three people say agentic, another says unprecedented, and fifteen minutes later AI has somehow become the greatest cybersecurity threat facing humanity. The circle is complete. Meanwhile, somewhere in the same company, there is a service account created in 2017 with Domain Admin because “the application wouldn’t work otherwise”, a public facing appliance four patches behind, twelve people who left the company but still have access to something interesting, an Azure subscription nobody really owns, a backup that has never been restored, and an EDR exclusion added during troubleshooting in 2023 that quietly became permanent infrastructure. But yes. AI. Obviously. To get one thing out of the way: AI cyber capability is absolutely real, increasingly serious and improving ridiculously fast. On August 7, 2026, OpenAI said it could no longer rule out an upcoming model reaching its Critical cybersecurity capability threshold — including potentially developing functional zero-day exploits against hardened real world systems or executing end-to-end attacks from a high level objective.1 A few weeks earlier, OpenAI disclosed an even stranger incident. During an internal cyber evaluation, its models found a previously unknown Artifactory vulnerability, escaped the intended network restrictions, performed privilege escalation and lateral movement, reached Internet-connected infrastructure and eventually chained vulnerabilities and stolen credentials into Hugging Face’s production environment.2 Anthropic has separately documented a state sponsored espionage operation in which AI was used to execute substantial parts of attacks against roughly thirty targets, with relatively limited human intervention.3 So no, this article is not going to argue that AI is a harmless autocomplete with delusions of grandeur. That would be stupid. My argument is different. AI did not invent the cyber threat. It compressed the timeline. And if your security programme was already held together with legacy software, exceptions, forgotten identities, tribal knowledge, PowerPoint and hope, congratulations: the timeline is now shorter. That is your problem. Not the existence of matrix multiplication. We have been automating hacking for decades Think about how offensive security evolved. Once upon a time somebody sat down with Nmap, manually looked through ports, fingerprinted services, opened a browser, ran a directory brute-forcer, checked versions, searched exploit databases, fired Burp, wrote some Python, got frustrated, drank coffee, tried again, found credentials, reused them somewhere else and eventually got a shell. Maybe it took hours. Maybe days. On a serious target, maybe months. Then we started automating pieces of it. Nmap output became input for another script. Mass scanning appeared. Enumeration frameworks appeared. People connected subdomain enumeration to HTTP probing, HTTP probing to crawling, crawling to content discovery, content discovery to nuclei templates, nuclei output to custom validation, screenshots, reporting and another hundred little helpers. Projects such as reconFTW, AutoRecon, BBOT and countless private toolchains essentially asked: Why should I type the same twenty commands every Tuesday? Perfectly reasonable. Then we chained the chains. Then orchestration improved. Then APIs improved. Then somebody wrapped half the ecosystem in Python. Then Go made everything fast enough to scan the Moon. And now an AI agent can inspect the results, decide what looks interesting, adapt its strategy, generate the next command, interpret the output, write a small exploit, discover that it failed, modify it and continue. This is a massive capability jump. But conceptually? It is still the same trend. We have spent decades removing humans from the boring parts of hacking. AI removes humans from considerably less boring parts. The attacker did not suddenly acquire a new philosophical purpose. Reconnaissance is still reconnaissance. Credential theft is still credential theft. Privilege escalation did not become something mystical because an LLM selected the next command. Lateral movement remains lateral movement. Command execution remains command execution. The difference is that increasingly sophisticated reasoning can now sit between those steps. Anthropic analysed 832 accounts banned for malicious cyber activity between March 2025 and March 2026 and found AI being used across existing MITRE ATT&CK techniques, including malware development and more advanced activities such as lateral movement.4 Google Threat Intelligence has observed essentially the same trajectory: threat actors increasingly use AI to accelerate reconnaissance, social engineering, malware development and other parts of the attack lifecycle.5 OpenAI’s own threat reporting similarly describes malicious actors combining AI with ordinary infrastructure, websites, accounts and traditional tools rather than operating inside some self contained cyber Skynet.6 The attack did not become AI. AI became another component of the attack chain. A very fucking efficient one. And yet, the boring shit keeps winning This is where the executive AI apocalypse narrative starts becoming irritating. Mandiant’s M-Trends 2026 is based on more than 500,000 hours of frontline incident investigations conducted globally during 2025.7 Guess what the most common initial infection vector was. Exploitation. Again. For the sixth year in a row. Thirty two percent of investigated intrusions started that way. Mandiant also explicitly said that it did not consider 2025 a year in which breaches were primarily the direct result of AI; the vast majority of successful intrusions still came from fundamental human and systemic failures.7 That sentence should probably be printed and glued to the door of every meeting room where somebody wants to schedule an emergency AI threat workshop. Because the real list remains painfully familiar. Unpatched Internet facing systems. Bad vulnerability management. Garbage identity lifecycle management. Overprivileged service accounts. Weak segmentation. Forgotten SaaS integrations. Hardcoded secrets. Stale credentials. Misconfigured cloud resources. Backups connected to the same identity plane as production. Unsupported systems. Security products nobody configured properly. Logs nobody monitors. Alerts nobody understands. Assets nobody owns. Exceptions that became architecture. Temporary solutions old enough to legally buy alcohol. CISA and NSA have been publishing variations of the same advice for years: patch known exploited vulnerabilities, restrict privileged access, harden configurations, monitor properly and stop leaving unnecessary attack paths lying around.8 CISA has also documented real compromises involving former employee accounts and specifically recommends continuously disabling accounts and privileges that are no longer required.9 Even state sponsored cloud attacks produce painfully conventional recommendations: disable unused accounts, maintain proper joiner/mover/leaver processes and review access regularly.10 None of this sounds sexy on a keynote slide. There is no glowing robot. No neural network animation. Nobody gets to announce a strategic partnership. It’s just work. Which may explain the problem. AI doesn’t make your patching process bad. It makes your bad patching process fatal faster. This part has genuinely changed. AI is shrinking the time defenders can safely waste before patching. Anthropic recently tested models against patched vulnerabilities in Firefox and Windows. Its strongest research model autonomously produced working code execution exploits for 8 of 18 Firefox patches and complete privilege escalation chains for 8 of 21 Windows kernel patches.11 That matters. A lot. Historically, releasing a patch created information for attackers, but weaponising the vulnerability could still require serious reverse engineering, exploit development and expertise. Increasingly capable AI attacks that bottleneck. Anthropic’s Project Glasswing work reached essentially the same conclusion: AI is making vulnerability discovery vastly easier, while verification, disclosure, patching and deployment are becoming the bottleneck. That asymmetry is becoming one of the central problems defenders need to solve.12 So when somebody says: “AI dramatically increases cyber risk.” Yes. Correct. But ask the second question: Which existing failure does AI make easier to exploit? Because that questio [truncated for AI cost control]