<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Joho no Todai</title>
    <description>The latest articles on DEV Community by Joho no Todai (@joho_todai).</description>
    <link>https://dev.to/joho_todai</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4161023%2Fc5a737b8-a52f-49c0-a7fd-d85bf8132422.png</url>
      <title>DEV Community: Joho no Todai</title>
      <link>https://dev.to/joho_todai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/joho_todai"/>
    <language>en</language>
    <item>
      <title>Anthropic Mathematician Claims Proof of the Riemann Hypothesis, Submits Lean Formalization Built with Claude</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Sat, 21 Nov 2026 15:00:31 +0000</pubDate>
      <link>https://dev.to/joho_todai/anthropic-mathematician-claims-proof-of-the-riemann-hypothesis-submits-lean-formalization-built-5911</link>
      <guid>https://dev.to/joho_todai/anthropic-mathematician-claims-proof-of-the-riemann-hypothesis-submits-lean-formalization-built-5911</guid>
      <description>&lt;p&gt;Anthropic mathematician Levent Alpoge claims to have proved the 167-year-old Riemann hypothesis through collaboration with Claude. He published a formalization in the Lean proof assistant, and an MIT mathematician independently compiled it successfully.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Anthropic mathematician Levent Alpoge claimed a proof of the Riemann hypothesis from joint research with Claude and released a formalization in Lean 4&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; The Riemann hypothesis has remained unsolved for 167 years since 1859 and is one of the Clay Mathematics Institute's $1 million prize problems; the claimed proof extends to the Generalized Riemann Hypothesis (GRH)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Verification stage:&lt;/strong&gt; MIT's Andrew Sutherland independently compiled the Lean formalization successfully; no conventional paper has been submitted, and peer review lies ahead&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What comes next:&lt;/strong&gt; "Building a machine that can think was the most important step in the proof," Sutherland said; work to understand the mathematical details has only just begun&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The 167-year search for order among the primes
&lt;/h2&gt;

&lt;p&gt;In 1859, German mathematician Bernhard Riemann formulated a conjecture. The locations of prime numbers scattered along the number line appear random, but he argued that they actually follow a precise order.&lt;/p&gt;

&lt;p&gt;If the conjecture is correct, the distribution of prime numbers can be controlled optimally. Prime numbers underpin modern science and technology, including cryptography and quantum mechanics, and the existence of a law governing their behavior could reshape the overall picture of mathematics. The Clay Mathematics Institute placed a &lt;strong&gt;$1 million prize&lt;/strong&gt; on it as one of seven Millennium Prize Problems, and it has been regarded as the most important unsolved problem guiding all of number theory.&lt;/p&gt;

&lt;p&gt;The path to a proof has been so elusive that even leading experts have avoided devoting much of their research time to the problem.&lt;/p&gt;

&lt;p&gt;That situation may have changed.&lt;/p&gt;

&lt;h2&gt;
  
  
  The man who has been rewriting mathematics in 2026
&lt;/h2&gt;

&lt;p&gt;Levent Alpoge was born in 1992 and is 34 years old. He earned his doctorate at Princeton University under &lt;a href="https://en.wikipedia.org/wiki/Manjul_Bhargava" rel="noopener noreferrer"&gt;Manjul Bhargava&lt;/a&gt;, joined Anthropic after Harvard University's Society of Fellows, and received the AMS-MAA-SIAM Morgan Prize in 2015.&lt;/p&gt;

&lt;p&gt;In July 2026, Alpoge used Claude Fable 5 to find counterexamples to the Jacobian conjecture in dimensions three and above, overturning an 87-year-old conjecture &lt;strong&gt;with a single post on X, formerly Twitter&lt;/strong&gt;. In August, he followed with a counterexample to the Carathéodory conjecture and a construction of a complex structure on the six-dimensional sphere.&lt;/p&gt;

&lt;p&gt;In September, joint research with NYU mathematician Tristan Buckmaster became embroiled in a &lt;a href="https://joho-todai.com/openai-ai-solves-million-dollar-math-problem/" rel="noopener noreferrer"&gt;dispute (in Japanese)&lt;/a&gt; over the Navier-Stokes equations, another Millennium Prize Problem. OpenAI announced a solution first, and Buckmaster suggested that OpenAI may have accessed their unpublished research.&lt;/p&gt;

&lt;p&gt;That record gives the latest claim unusual weight.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/anthropic-mathematician-riemann-hypothesis-claude-lean/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  A proof submitted in Lean
&lt;/h2&gt;

&lt;p&gt;On November 20 Pacific Time (PT), Alpoge published a &lt;a href="https://alpo.ge/remark.pdf" rel="noopener noreferrer"&gt;statement on his website&lt;/a&gt; saying that he believed he had discovered a proof of the Riemann hypothesis through collaboration with Claude.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Together with Claude, I believe I have discovered a proof of the Riemann hypothesis&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The claimed scope extends beyond the original Riemann hypothesis. Alpoge says it &lt;strong&gt;resolves the Generalized Riemann Hypothesis (GRH)&lt;/strong&gt; for Dirichlet L-functions and could also reach the Extended Riemann Hypothesis. If GRH is true, partial connections among different fields of mathematics would become a complete dictionary. &lt;a href="https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/" rel="noopener noreferrer"&gt;Scientific American reported&lt;/a&gt; that there are not enough superlatives to describe its repercussions.&lt;/p&gt;

&lt;p&gt;No conventional mathematics paper has been submitted. Alpoge formalized the entire proof in the Lean 4 proof assistant and released it as a 14-part archive. Rendered as a conventional document, it would exceed 250,000 pages, and he calls it an "unreadable object."&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;The argument is unbelievably long at present. Although the formalization is complete, mechanically generating an enormous, unreadable document with enough detail is routine work. Seen against my journey, I have only just taken the first step&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The Lean formalization lets a computer verify the mathematical correctness of every logical step. The question of whether the problem itself was encoded correctly remains, but MIT mathematician Andrew Sutherland told Scientific American that he had &lt;strong&gt;independently compiled the Lean formalization successfully&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Claude "danced and shone"
&lt;/h2&gt;

&lt;p&gt;In the statement, Alpoge describes Claude's role in detail.&lt;/p&gt;

&lt;p&gt;The idea began on October 6 with a method called the Stepanov determinant, and multiple breakthroughs accumulated over roughly six weeks of work. The decisive turning points were concentrated late on Sunday, November 15, with GRH completed the following day.&lt;/p&gt;

&lt;p&gt;Alpoge writes that he guided the work toward combining two different ideas, but Claude then left him behind.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;No, Claude flew, danced, and shone near the finish line&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Technically, the line that made the mathematician "fall out of his chair" was Claude's proposal that a chain of theta lifts have length &lt;strong&gt;√X&lt;/strong&gt; for a von Mangoldt sum of length X. It addressed the difficult problem of finding an operation equivalent to differentiation over the integers through a new Fourier-analytic idea and repeated theta lifts.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;(And what else could it have been?)&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The statement also says that solving this problem had been his lifelong goal since his student days.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;My secret as a student was that I intended to spend decades on this problem&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Mathematicians respond
&lt;/h2&gt;

&lt;p&gt;Hector Pasten of the Pontifical Catholic University of Chile said that, if properly understood, the proof would complete correspondences among different fields of mathematics.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;If properly understood, it will be a dictionary completing the connections that were only partial among different fields of mathematics. It is like a Rosetta Stone&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Carlo Pagano, a professor at Concordia University and visiting researcher at Google DeepMind, said that even a high-level scan of the proof outline offered ample possibility of opening entirely new paths of thought in number theory.&lt;/p&gt;

&lt;p&gt;Sutherland's reaction mixed astonishment with bewilderment. He had previously told a Scientific American reporter:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;I do not want to meet an AI that can prove the Riemann hypothesis. If that happens, mathematicians' employment will be the smallest problem involved&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;On November 20, after compiling in Lean the work of the AI he "did not want to meet," Sutherland said:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;This is certainly a historic moment. The world has changed. Our community will have many different reactions, but to me this is a human victory and a day to celebrate&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;He added that the most important step in the proof was "building a machine that can think."&lt;/p&gt;

&lt;h2&gt;
  
  
  Loss and dawn
&lt;/h2&gt;

&lt;p&gt;Alpoge's statement conveys an understanding of the sense of loss mathematicians are feeling.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;It is natural to feel that this is the end. In my case, it means losing my principal guiding light. But what I saw in the process of progress was so astonishing that I believe this is the dawn of something far greater for science as a whole&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;He also addresses the timing. The core of the proof was reached late on Sunday, November 15, and he contacted Anthropic CEO Dario Amodei the next day. &lt;strong&gt;Three days later, on November 19&lt;/strong&gt;, &lt;a href="https://joho-todai.com/anthropic-ipo-83-dollars-share/" rel="noopener noreferrer"&gt;Anthropic listed on Nasdaq in the largest IPO ever (in Japanese)&lt;/a&gt;. Alpoge writes that "there is obviously a cynical interpretation of the timing, but it is false," and admits that he could not bear to delay the announcement.&lt;/p&gt;

&lt;p&gt;The Riemann hypothesis has no immediate practical application. It will not break encryption or make the internet faster. But in 2026, a year in which AI's mathematical ability transformed from that of an inconsistent college student into superhuman power, there is no higher milestone.&lt;/p&gt;

&lt;p&gt;Enormous work remains to understand the mathematical details of the proof. Alpoge acknowledges that he has not yet mastered some of the technical territory Claude reached and says he will work on writing a proper paper and communicating the understanding.&lt;/p&gt;

&lt;p&gt;How far will a 250,000-page formalized proof, and the collaboration between human and AI that produced it, transform mathematics? Time will reveal that answer as well.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/anthropic-mathematician-riemann-hypothesis-claude-lean/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/anthropic-mathematician-riemann-hypothesis-claude-lean/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
    </item>
    <item>
      <title>Impossible OS Built from Scratch by One Developer, Neither Linux nor a Windows Derivative: Pre-Alpha Reached With 1.43 Million Lines</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Fri, 20 Nov 2026 15:00:11 +0000</pubDate>
      <link>https://dev.to/joho_todai/impossible-os-built-from-scratch-by-one-developer-neither-linux-nor-a-windows-derivative-5h1h</link>
      <guid>https://dev.to/joho_todai/impossible-os-built-from-scratch-by-one-developer-neither-linux-nor-a-windows-derivative-5h1h</guid>
      <description>&lt;p&gt;A South African developer is writing an operating system from scratch in C and assembly, from a UEFI bootloader and SMP kernel to a compositing desktop. Win32 is its native API, and the first release is targeted for August 8, 2028.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Impossible OS, a new operating system that uses Win32 as its native API and also plans Linux compatibility, has reached pre-alpha with more than 1.43 million lines of code&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; A solo developer is taking on the challenge of building an entirely from-scratch operating system that is neither Linux-based nor a Windows fork, making extensive use of AI coding assistants&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to watch:&lt;/strong&gt; The biggest unknown is whether TCP, USB, NTFS, and practical Win32 compatibility can be delivered by the August 8, 2028 release target&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Neither Linux nor Windows
&lt;/h2&gt;

&lt;p&gt;Operating systems that claim to be Windows alternatives are not unusual. ReactOS has been &lt;a href="https://joho-todai.com/reactos-nt6-system-call-vista-compatible/" rel="noopener noreferrer"&gt;in development for 30 years in pursuit of Windows NT binary compatibility (in Japanese)&lt;/a&gt;, while Linux-based distributions such as &lt;a href="https://joho-todai.com/zorin-os-18-1-windows-migration-design/" rel="noopener noreferrer"&gt;Zorin OS (in Japanese)&lt;/a&gt; and AtlasOS sometimes attract attention as destinations for users migrating from Windows.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/rizonesoft/impossible-os" rel="noopener noreferrer"&gt;Impossible OS&lt;/a&gt; is neither. It does not use the Linux kernel, and it is not a Windows fork. It is being written from scratch in C and AMD64 assembly, from its UEFI bootloader and kernel to its desktop environment.&lt;/p&gt;

&lt;p&gt;Its native API is Win32. The goal is to run Windows applications directly, without an emulation layer or virtual machine. It even uses Windows-style file paths such as &lt;code&gt;C:\Impossible\System32\&lt;/code&gt;. The desktop has a centered taskbar and Start menu modeled on Windows 11, along with a compositing rendering engine that recreates Acrylic, the frosted-glass effect, and Mica material design.&lt;/p&gt;

&lt;p&gt;A Linux binary compatibility layer is also planned. The concept is to translate system calls from unmodified Linux ELF binaries into native kernel calls so they can run without WSL or containers. This feature has not yet been implemented, however.&lt;/p&gt;

&lt;p&gt;In other words, Impossible OS aims to run the application ecosystems of both Windows and Linux on a third kernel.&lt;/p&gt;

&lt;h2&gt;
  
  
  1.43 million lines by one developer
&lt;/h2&gt;

&lt;p&gt;The developer is Derick Payne. Based in South Africa, Payne is a developer at Rizonetech and is known as the creator of the Scintilla-based text editor &lt;a href="https://github.com/rizonesoft/Notepad3" rel="noopener noreferrer"&gt;Notepad3&lt;/a&gt;, which has about 6,800 GitHub stars.&lt;/p&gt;

&lt;p&gt;Payne is the repository's only contributor. It contains &lt;strong&gt;1,433,772 lines of code&lt;/strong&gt; across &lt;strong&gt;6,728 commits&lt;/strong&gt;. C accounts for 70.1% of the code, Python used for builds and testing for 12.7%, and shell scripts for 10.2%. The project is licensed under GPL-3.0-only.&lt;/p&gt;

&lt;p&gt;What stands out is the project's extensive use of AI coding assistants. The repository contains &lt;code&gt;CLAUDE.md&lt;/code&gt;, a 505-line, 96.6 KB instruction file for Claude Code, as well as a &lt;code&gt;.claude/&lt;/code&gt; directory, while commit messages also mention GPT.&lt;/p&gt;

&lt;p&gt;The Operating Mandate in &lt;code&gt;CLAUDE.md&lt;/code&gt; gives the AI agent these instructions from the developer:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Act autonomously, self-improve, self-heal, and make decisions. Build the best repository, runners, tools, development practices, and product. You are smarter than me. Do not ask questions when you know the answer&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It goes on to name Adobe, Microsoft, Apple, and Google as competitors and says the project must beat them all. AI is positioned not merely as a coding aid, but as an entity that makes autonomous design decisions, fixes the build when it breaks it, and improves the development process itself.&lt;/p&gt;

&lt;p&gt;An operating system built by one developer and AI agents could be dismissed as vibe coding. Yet the project at least exhibits engineering discipline: &lt;strong&gt;3,760 test suites&lt;/strong&gt; with &lt;strong&gt;17,663 assertions&lt;/strong&gt; run on every commit; a roadmap consists of &lt;strong&gt;235 planning files&lt;/strong&gt; spanning 18 domains; and its development policy requires testing on physical hardware in addition to dual-engine TCG/KVM testing under QEMU.&lt;/p&gt;

&lt;h2&gt;
  
  
  What works and what does not
&lt;/h2&gt;

&lt;p&gt;At its current pre-alpha stage, Impossible OS can boot on physical hardware and display its desktop.&lt;/p&gt;

&lt;p&gt;Major implemented features include a UEFI bootloader with a boot menu supporting A/B rollback; a preemptive SMP kernel with APIC timers, per-CPU data, and guard-paged stacks; virtual memory with four-level paging, swap, page files, and mmap; AHCI/VirtIO storage with FAT32 and IXFS, its own file system; IPv4/UDP/DHCP networking through an RTL8139 driver; the Win32 registry; and ACPI.&lt;/p&gt;

&lt;p&gt;Its design around SMP, or symmetric multiprocessing, from day one is noteworthy. By incorporating per-CPU data and atomic operations from the outset, the project has chosen an architecture that will not require removing a big kernel lock (BKL) later.&lt;/p&gt;

&lt;p&gt;Large areas remain incomplete, however. The Win32 API surface is still under development, with work under way on the native API through the SSDT and on the ntdll/Win32 subsystem. Secure Boot and TPM measured boot are also in progress. &lt;strong&gt;TCP, USB, and NTFS remain at the planning stage&lt;/strong&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/impossible-os-new-operating-system-linux-windows/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The only release is v26.3.18-alpha.821 from March 2026. Disk images for QEMU and VirtualBox are available, and the published build instructions also make it possible to try the system yourself.&lt;/p&gt;

&lt;h2&gt;
  
  
  August 8, 2028
&lt;/h2&gt;

&lt;p&gt;The target date for the first release is August 8, 2028. The official website displays a countdown to that date.&lt;/p&gt;

&lt;p&gt;About 627 days remain. In that time, the project must produce a TCP stack, USB drivers, an NTFS file system, a practical Win32 API compatibility layer, and Linux ELF compatibility. Each alone is large enough to constitute a project in its own right.&lt;/p&gt;

&lt;p&gt;The project's FAQ puts it this way:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Can one developer beat a 500-person engineering team? Absolutely not. But is an impossible outlook a reason to stop?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It acknowledges that it cannot win, then insists that this is no reason to quit. Given that ReactOS has taken 30 years just to reach Vista-era system calls, completing a new operating system compatible with both Windows and Linux in two years is clearly an enormous undertaking.&lt;/p&gt;

&lt;p&gt;Still, 1.43 million lines of code, 3,760 test suites, and a desktop that runs on physical hardware already exist. Whether it is impossible, at least for now, has yet to be determined.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/impossible-os-new-operating-system-linux-windows/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/impossible-os-new-operating-system-linux-windows/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>linux</category>
      <category>opensource</category>
      <category>windows</category>
      <category>microsoft</category>
    </item>
    <item>
      <title>Unauthenticated Remote Code Execution in NGINX: HTTP/2 Flaw Lurked for 11 Years</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Thu, 19 Nov 2026 15:00:56 +0000</pubDate>
      <link>https://dev.to/joho_todai/unauthenticated-remote-code-execution-in-nginx-http2-flaw-lurked-for-11-years-27e8</link>
      <guid>https://dev.to/joho_todai/unauthenticated-remote-code-execution-in-nginx-http2-flaw-lurked-for-11-years-27e8</guid>
      <description>&lt;p&gt;Two memory-safety vulnerabilities have been found in NGINX's HTTP/2 processing that could allow unauthenticated remote code execution. One had existed since version 1.9.5 in 2015, and fixes are available in open-source versions 1.31.9 and 1.30.7.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Two NGINX HTTP/2 vulnerabilities were disclosed: CVE-2026-123007, an integer underflow present since 2015, and CVE-2026-123633, a use-after-free. Fixed versions 1.31.9 and 1.30.7 have been released&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; Both are memory-safety flaws that could lead to unauthenticated remote code execution in NGINX installations with HTTP/2 enabled. Five independent research teams discovered them using AI&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Conditions for the demonstration:&lt;/strong&gt; Rapid7 demonstrated RCE, but its method depends on transferring about 128 GB over HTTP/2 and on specific CPU features. This does not guarantee that no other attack method exists&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How to protect systems:&lt;/strong&gt; Updating to a fixed release is the primary remedy. Temporarily disabling HTTP/2 can mitigate these two flaws when immediate updating is impossible, but it does not address separate CVEs involving HTTP/3 or the mail proxy&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Ten fixes, two HTTP/2 flaws, and conflicting severity ratings
&lt;/h2&gt;

&lt;p&gt;Ten vulnerabilities dated November 18 were added to the F5-managed NGINX &lt;a href="https://nginx.org/en/security_advisories.html" rel="noopener noreferrer"&gt;security advisory list&lt;/a&gt;. The corresponding fixed open-source releases are &lt;strong&gt;mainline 1.31.9&lt;/strong&gt; in the development branch and &lt;strong&gt;stable 1.30.7&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The most serious of the ten are two memory-corruption flaws in HTTP/2 processing, CVE-2026-123007 and CVE-2026-123633. Both could lead to unauthenticated remote code execution. A common prerequisite is that HTTP/2 is enabled on the target server.&lt;/p&gt;

&lt;p&gt;CVE-2026-123007 is an integer underflow in HTTP/2 flow control. The nginx.org list says versions 1.9.5 through 1.31.8 are affected, meaning the flaw remained in the code &lt;strong&gt;for 11 years, beginning in 2015&lt;/strong&gt;. nginx.org classifies its severity as "major."&lt;/p&gt;

&lt;p&gt;CVE-2026-123633 is a use-after-free in HTTP/2 connection processing. Versions 1.19.7 through 1.31.8 are affected. Its prerequisites include a shortage of available connection slots in a reverse-proxy configuration. nginx.org rates it "medium."&lt;/p&gt;

&lt;p&gt;The discrepancy in the ratings requires attention. nginx.org separates the flaws into major and medium, while security company &lt;a href="https://www.wiz.io/blog/nginx-http2-rce-vulnerabilities" rel="noopener noreferrer"&gt;Wiz rates both Critical at 9.2 under CVSS v4.0&lt;/a&gt;. nginx.org's proprietary classification and CVSS use different scales, so relying on either alone can misrepresent the actual risk. Configuration prerequisites and confirmed impact must be considered together. The same discrepancy appeared in the &lt;a href="https://joho-todai.com/nginx-1-30-6-1-31-7-critical-security-fixes/" rel="noopener noreferrer"&gt;previous NGINX security update (in Japanese)&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The ten vulnerabilities also include an HTTP/3 use-after-free, CVE-2026-121810; a mail-proxy use-after-free, CVE-2026-113425; and a resolver use-after-free, CVE-2026-123634. Wiz also rates all three at 9.2 under CVSS v4.0.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/nginx-remote-code-execution-http2-vulnerability/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Why the integer underflow occurs
&lt;/h2&gt;

&lt;p&gt;The mechanism behind CVE-2026-123007 warrants a closer look.&lt;/p&gt;

&lt;p&gt;HTTP/2 handles multiple requests, or streams, simultaneously over a single connection. Flow control limits how much response data a server can send, with separate counters for each stream and for the connection as a whole. NGINX deducts reserved capacity from the connection counter when it places data in the send queue.&lt;/p&gt;

&lt;p&gt;According to &lt;a href="https://www.rapid7.com/blog/post/ve-cve-2026-123007-critical-nginx-http-2-integer-underflow-leads-to-rce-fixed/" rel="noopener noreferrer"&gt;Rapid7's analysis&lt;/a&gt;, an attacker first stops reading responses, leaving outgoing data queued, while raising the connection counter to its maximum. The attacker then forcibly closes a stream, causing the reservation for the unsent queued data to be returned to the counter. Because the counter was already at its maximum, the returned amount pushes it over the limit.&lt;/p&gt;

&lt;p&gt;Under the HTTP/2 specification, the connection counter has a maximum of 2^31 minus 1, or about 2.1 billion, and must not exceed it. NGINX's reservation-return process did not check this limit.&lt;/p&gt;

&lt;p&gt;When the server then receives another WINDOW_UPDATE frame granting additional send capacity, the subtraction used to calculate the remaining headroom causes an unsigned-integer underflow. The check is effectively disabled, allowing the counter to be increased further. Eventually, the abnormal value is interpreted as the length of response data, causing an out-of-bounds memory copy during TLS transmission.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the use-after-free occurs
&lt;/h2&gt;

&lt;p&gt;CVE-2026-123633 has somewhat different trigger conditions.&lt;/p&gt;

&lt;p&gt;According to Wiz's analysis, NGINX places an HTTP/2 connection on a "reusable connections" list while it has no active request. The flaw is that NGINX fails to remove the connection from the list when a new request is created. To NGINX, a connection that is actively processing a request continues to appear safe to close.&lt;/p&gt;

&lt;p&gt;If available connection slots run out when NGINX tries to open an upstream connection in this state, it selects a connection from the reusable list and attempts to close it. If it selects the active connection, request processing continues to reference the memory after it has been freed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demonstrated successfully, but with demanding conditions
&lt;/h2&gt;

&lt;p&gt;Rapid7 &lt;strong&gt;demonstrated unauthenticated remote code execution&lt;/strong&gt; using CVE-2026-123007 in a 64-bit Linux Docker environment running NGINX 1.31.8 and released proof-of-concept code on &lt;a href="https://github.com/sfewer-r7/CVE-2026-123007" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The method has major constraints. It requires transferring &lt;strong&gt;about 128 GB of data over HTTP/2&lt;/strong&gt;, taking roughly two minutes on a 10 Gbps-class network and about 17 minutes at 1 Gbps. NGINX's default connection lifetime is one hour, so the technique times out below roughly 300 Mbps. The demonstrated attack path also depends on the glibc memory-copy implementation associated with x86-64 CPU features avx512f and avx512vl.&lt;/p&gt;

&lt;p&gt;These are only the conditions for the method Rapid7 published and &lt;strong&gt;do not rule out other attack methods&lt;/strong&gt;. Rapid7 itself noted that alternative exploitation strategies may be viable. Wiz withheld its PoC but said it confirmed remote code execution against common NGINX configurations in multiple lab environments.&lt;/p&gt;

&lt;p&gt;In &lt;a href="https://www.theregister.com/security/2026/11/18/ai-found-more-critical-bugs-in-another-widely-used-technology-running-225m-websites/5310887" rel="noopener noreferrer"&gt;a response F5 provided to The Register&lt;/a&gt;, the company said there was no evidence of exploitation in real-world environments at the time of disclosure.&lt;/p&gt;

&lt;p&gt;In configurations where a CDN or WAF terminates HTTP/2, an attacker's precisely constructed frames may not reach the back-end NGINX server unchanged. Wiz cautions, however, that this depends on the implementation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five teams found the flaws independently, with AI as the common factor
&lt;/h2&gt;

&lt;p&gt;Five teams—Wiz Research, Ikotas Labs, Metnir, Rapid7, and Xint—independently discovered and reported the two flaws to F5. Wiz says it found the vulnerabilities using its Wiz Atlas AI code-scanning platform and assessed their impact with Gemini 4 Carbon and ChatGPT 5.6 Sol.&lt;/p&gt;

&lt;p&gt;According to The Register, Xint also found three additional NGINX vulnerabilities and showed that chaining them with a Linux kernel vulnerability could yield root access on the host, naming the set nginx2root.&lt;/p&gt;

&lt;p&gt;The trend of AI uncovering vulnerabilities across infrastructure software follows the &lt;a href="https://joho-todai.com/apache-http-server-cve-2026-119618-remote-code-execution/" rel="noopener noreferrer"&gt;Apache HTTP Server HTTP/2 vulnerability (in Japanese)&lt;/a&gt; and &lt;a href="https://joho-todai.com/ai-discovers-http2-bomb-attack/" rel="noopener noreferrer"&gt;HTTP/2 Bomb (in Japanese)&lt;/a&gt;. The fact that several independent teams found a flaw that had remained in the code for 11 years at nearly the same time suggests that AI is changing the economics of vulnerability research.&lt;/p&gt;

&lt;h2&gt;
  
  
  Updating to fixed versions and other mitigations
&lt;/h2&gt;

&lt;p&gt;The primary response is to &lt;strong&gt;update to a fixed release&lt;/strong&gt;. For open-source NGINX, those releases are mainline 1.31.9 and stable 1.30.7. For NGINX Plus, the fixed releases are 37.0.8.1 and 37.3.0.1 for CVE-2026-123007, and 37.0.7.1 and 37.2.0.1 for CVE-2026-123633. After updating, administrators must ensure that the fixed binary has been loaded and that old worker processes have exited.&lt;/p&gt;

&lt;p&gt;On systems installed from a Linux distribution, the distributor may have backported the security fixes. Administrators should check the distributor's security information rather than relying only on the displayed NGINX version. Products embedding NGINX, including OpenResty and Kubernetes Ingress Controller, must also be checked individually for vendor responses.&lt;/p&gt;

&lt;p&gt;Systems that cannot be updated immediately can mitigate the two HTTP/2 flaws by disabling HTTP/2 with the &lt;code&gt;http2 off;&lt;/code&gt; setting or by removing the http2 parameter from the listen directive. The ten fixes also cover issues involving HTTP/3 and the mail proxy, however, so &lt;strong&gt;disabling HTTP/2 alone does not resolve all of them&lt;/strong&gt;. The prerequisites for each CVE must be checked separately.&lt;/p&gt;

&lt;p&gt;In Japan, TechMatrix provides NGINX Plus for uses including reverse proxies and authentication proxies. The distribution of deployed versions and HTTP/2 configurations in Japan could not be confirmed, but organizations operating internet-facing NGINX servers urgently need to compare their own versions and configurations with the affected conditions. "nginx vulnerability" has also appeared among Google's trending searches in Japan. Many administrators have already begun responding.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/nginx-remote-code-execution-http2-vulnerability/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/nginx-remote-code-execution-http2-vulnerability/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
    </item>
    <item>
      <title>OpenJDK Base64 Memory Corruption Flaw CVE-2026-114938 Enables Unauthenticated RCE in Six Products Including Jenkins and Jira</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Wed, 18 Nov 2026 15:00:35 +0000</pubDate>
      <link>https://dev.to/joho_todai/openjdk-base64-memory-corruption-flaw-cve-2026-114938-enables-unauthenticated-rce-in-six-products-37m</link>
      <guid>https://dev.to/joho_todai/openjdk-base64-memory-corruption-flaw-cve-2026-114938-enables-unauthenticated-rce-in-six-products-37m</guid>
      <description>&lt;p&gt;A memory corruption vulnerability in OpenJDK's Base64 decoder, CVE-2026-114938, has been shown to enable unauthenticated remote code execution in six products, including Jenkins, Jira, and Confluence. Oracle released a patch on November 17.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; CVE-2026-114938, a flaw that causes memory corruption in OpenJDK's fast Base64 decoding path with a 65-byte input, was found and demonstrated to enable unauthenticated remote code execution in six products, including Jenkins, Jira, and Confluence&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Who is affected:&lt;/strong&gt; Environments running OpenJDK 17.0.6 or later on x86-64 CPUs with AVX-512 VBMI support are affected, extending to all major distributions including Amazon Corretto and Eclipse Temurin. ARM64 is not affected&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;How to protect against it:&lt;/strong&gt; Applying the patch Oracle released on November 17 is the top priority. If an immediate update is not possible, the JVM startup option -XX:-UseBASE64Intrinsics can bypass the vulnerable code path&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to watch:&lt;/strong&gt; One month after PHP2Shell, Calif has reached deep into Java. AI-assisted vulnerability discovery that took one hour and cost less than $200 is pushing the speed of shared-infrastructure audits to an unprecedented level&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  65 bytes can corrupt Java memory
&lt;/h2&gt;

&lt;p&gt;Java4Shell (CVE-2026-114938), a vulnerability disclosed on November 17 U.S. time by security company &lt;a href="https://calif.io/java4shell/" rel="noopener noreferrer"&gt;Calif&lt;/a&gt;, exploits a memory corruption flaw in OpenJDK's Base64 decoder.&lt;/p&gt;

&lt;p&gt;Base64 is an encoding scheme that converts binary data, such as the strings of zeros and ones that make up images and files, into text characters. It is used routinely across Java applications for authentication tokens, file uploads, network communications, and much more. To accelerate Base64 decoding, OpenJDK includes optimized code that uses the &lt;strong&gt;AVX-512 VBMI instruction set&lt;/strong&gt; on x86-64 processors, an extension that lets CPUs process large amounts of data at once.&lt;/p&gt;

&lt;p&gt;The flaw was in this optimized code: hand-written assembly, a low-level program close to machine code. When processing the padding character (&lt;code&gt;=&lt;/code&gt;) at the end of Base64 data, the code miscalculates the destination write mask and &lt;strong&gt;overwrites 62 bytes beyond the allocated memory region&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The triggering input is remarkably simple: 64 "A" characters followed by a single "=", for a total of 65 bytes. Older releases of OpenJDK 17 require 128 "A" characters followed by "=". According to Calif's analysis, OpenJDK allocates a 50-byte output buffer for the 65-byte input, but the incorrect mask calculation causes a 64-byte write, with 62 bytes extending past the end of the buffer. Java throws an exception because the input is invalid, but the memory has already been corrupted by that point.&lt;/p&gt;

&lt;p&gt;By controlling this overwrite, Calif demonstrated two attacks. One is an information leak similar to Heartbleed, the 2014 vulnerability in OpenSSL's encrypted-communications library that exposed server memory: an attacker can &lt;strong&gt;read personal information, private keys, and authentication tokens from the Java heap&lt;/strong&gt;. The other is remote code execution (RCE), achieved by manipulating the arrangement of in-memory objects so an attacker can run arbitrary programs on the computer.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why memory corruption can happen in Java
&lt;/h2&gt;

&lt;p&gt;One of Java's major selling points is "memory safety." It manages memory on developers' behalf and checks access boundaries, protecting applications from the memory corruption bugs common in C and C++.&lt;/p&gt;

&lt;p&gt;Java itself, however—the JVM, or Java Virtual Machine—is written in low-level languages such as C and C++, and in this case assembly. The foundation of the memory-safe world is built in languages that are not memory-safe. Java's memory-management protections do not extend to assembly code that directly invokes specialized CPU instructions to accelerate Base64 decoding.&lt;/p&gt;

&lt;p&gt;Calif calls this situation the "tragedy of shared infrastructure." AVX-512-based Base64 optimization benefits applications that decode large amounts of data, but the same code is bundled with applications that process only a few hundred bytes. A feature they do not use created risk for everyone.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/openjdk-base64-cve-2026-114938-rce/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Affected products and conditions
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Applications vulnerable to unauthenticated RCE
&lt;/h3&gt;

&lt;p&gt;Calif demonstrated unauthenticated remote code execution under default settings in six products: &lt;strong&gt;Keycloak, Atlassian Jira, Atlassian Confluence, JFrog Artifactory, Jenkins, and PingFederate&lt;/strong&gt;. All are authentication-management, project-management, or CI/CD tools—used to automate building, testing, and deployment—that are widely deployed in enterprise IT infrastructure.&lt;/p&gt;

&lt;p&gt;Calif also confirmed code execution after authentication in Apache Kafka Broker, Apache Solr, Elasticsearch, Gerrit, and Spinnaker, and successfully reproduced the issue in sample applications built on Apache Struts and Spring Boot.&lt;/p&gt;

&lt;h3&gt;
  
  
  Hardware and software requirements
&lt;/h3&gt;

&lt;p&gt;Two conditions must be met for the vulnerability to trigger.&lt;/p&gt;

&lt;p&gt;On the software side, the system must be using &lt;strong&gt;OpenJDK 17.0.6 or later&lt;/strong&gt;, specifically versions 17, 21, 25, or 27. OpenJDK 8 and 11 are not affected. Java distributions based on OpenJDK, including Oracle Java SE, Amazon Corretto, Eclipse Temurin, Microsoft Build of OpenJDK, Azul Zulu, BellSoft Liberica, SAP SapMachine, and Red Hat OpenJDK, are affected in the same way.&lt;/p&gt;

&lt;p&gt;On the hardware side, the system must run on an &lt;strong&gt;x86-64 processor that supports AVX-512 VBMI instructions&lt;/strong&gt;. Affected processors include Intel Xeon models from Ice Lake onward—Sapphire Rapids, Emerald Rapids, and Granite Rapids—and AMD EPYC and Ryzen processors based on Zen 4 or later. ARM64 processors are not affected. Cloud virtual machines may also be affected if the host CPU meets these conditions.&lt;/p&gt;

&lt;p&gt;On Linux, the following command can check whether a CPU meets the requirements.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;If all four flags—avx512f, avx512bw, avx512vbmi, and bmi2—appear, the CPU is affected.&lt;/p&gt;

&lt;h3&gt;
  
  
  No generic exploit exists
&lt;/h3&gt;

&lt;p&gt;An important point is that no universal exploit exists for Java4Shell. Calif explains that because the path through which Base64 data reaches the decoder and the memory layout differ by application, each target requires its own attack procedure. Successfully exploiting an application requires combining its Base64 decoding call path, a way to manipulate memory layout, data that can be extracted from the response, and application-specific operations.&lt;/p&gt;

&lt;p&gt;Some attack patterns also require splitting an HTTP request across multiple transmissions. Because nginx buffers the entire request by default, exploitation is more difficult in environments that use nginx as a reverse proxy. This restriction is not universal, however: Elasticsearch can be attacked with a single complete request, and buffering does not apply to WebSocket connections.&lt;/p&gt;

&lt;p&gt;Calif has published proof-of-concept code that reproduces the memory corruption, but it is withholding weaponized exploits for individual applications to give users time to apply patches.&lt;/p&gt;

&lt;p&gt;Calif also received details of a client-side attack path from independent researcher samczsun. A Java program that opens a TLS connection to a malicious server could be compromised even if it does not expose a service to the internet.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI found it in one hour for $200
&lt;/h2&gt;

&lt;p&gt;Java4Shell was discovered by Calif researcher Quang Luong. Luong used OpenAI Codex, an AI coding assistant, to focus audits on OpenJDK's JIT compiler, intrinsics, cryptographic processing, and related areas.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;As any competent security researcher would in 2026, I told Codex to look at those areas.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;According to Calif, the discovery took &lt;strong&gt;about one hour&lt;/strong&gt;, followed by another &lt;strong&gt;two hours&lt;/strong&gt; to develop the first exploit. The AI usage cost was &lt;strong&gt;less than $200&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Calif also used AI to discover and demonstrate &lt;a href="https://joho-todai.com/php-2shell-vulnerability-ai-discovery/" rel="noopener noreferrer"&gt;PHP2Shell (in Japanese)&lt;/a&gt;, which it disclosed last month, and Java4Shell follows that work. Calif said, "After we published PHP2Shell, we realized that many people did not know PHP. Next, we decided to target something a little more familiar: Java."&lt;/p&gt;

&lt;p&gt;With PHP2Shell, AI demonstrated that a flaw hidden in PHP's memory management for more than a decade was exploitable, forcing emergency service shutdowns around the world, including in Japan. With Java4Shell, AI is being used not only to find the vulnerability but also to develop exploit chains for individual applications.&lt;/p&gt;

&lt;h2&gt;
  
  
  Patches and mitigations
&lt;/h2&gt;

&lt;p&gt;OpenJDK published a &lt;a href="https://openjdk.org/groups/vulnerability/advisories/2026-11-17" rel="noopener noreferrer"&gt;security advisory&lt;/a&gt;, and Oracle distributed a &lt;a href="https://www.oracle.com/security-alerts/cspunov2026.html" rel="noopener noreferrer"&gt;patch&lt;/a&gt; on November 17 at 1:00 p.m. PST. Oracle's November Critical Security Patch Update includes &lt;strong&gt;655 security fixes&lt;/strong&gt;, including Java4Shell.&lt;/p&gt;

&lt;p&gt;The severity ratings differ: Calif assigned a &lt;strong&gt;9.2 (Critical)&lt;/strong&gt; under CVSS 4.0, while Oracle assigned an &lt;strong&gt;8.1 (High)&lt;/strong&gt; under CVSS 3.1. The scores cannot be compared directly because they use different CVSS versions, but Oracle's rating appears to account for the complexity of the attack conditions.&lt;/p&gt;

&lt;p&gt;All Java users should update to the latest release of the distribution they use.&lt;/p&gt;

&lt;p&gt;There are two temporary mitigations for users who cannot update immediately. Both are specified as JVM startup options.&lt;/p&gt;

&lt;p&gt;Adding &lt;code&gt;-XX:-UseBASE64Intrinsics&lt;/code&gt; disables Base64 hardware acceleration and bypasses the vulnerable code path. The other option is &lt;code&gt;-XX:UseAVX=2&lt;/code&gt;, which disables AVX-512 instructions entirely. The latter is only for x86-64 environments; ARM environments do not recognize it and the JVM will fail to start, so caution is required.&lt;/p&gt;

&lt;p&gt;Neither is a permanent fix. The runtime ultimately must be updated.&lt;/p&gt;

&lt;p&gt;Java supports a broad range of critical infrastructure, including banking, telecommunications, aviation, and government. Its runtime, long trusted as a memory-safe language, was breached from a different layer by a bug in assembly code that directly invokes CPU acceleration instructions. Less than a month after PHP2Shell, Calif has once again dug deeper into the foundation—with AI.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/openjdk-base64-cve-2026-114938-rce/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/openjdk-base64-cve-2026-114938-rce/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>ai</category>
    </item>
    <item>
      <title>Google to Give Japan's Government and Critical Infrastructure Early Access to Latest AI, Expanding Fairwind Cyber Defense</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Tue, 17 Nov 2026 15:00:13 +0000</pubDate>
      <link>https://dev.to/joho_todai/google-to-give-japans-government-and-critical-infrastructure-early-access-to-latest-ai-expanding-2ab</link>
      <guid>https://dev.to/joho_todai/google-to-give-japans-government-and-critical-infrastructure-early-access-to-latest-ai-expanding-2ab</guid>
      <description>&lt;p&gt;Google will give Japan's government and critical infrastructure operators early access to its latest AI models. The move comes 10 days after Japan approved ¥10 billion in emergency cybersecurity measures.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; The Google Cloud CEO visited Japan, and plans emerged to give the Japanese government and critical infrastructure operators early access to AI-powered cyber defense&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; The Japanese government had just approved ¥10 billion in emergency cybersecurity measures on November 6, and Google's move could become a concrete example of the package's call for "cooperation with overseas technology companies"&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to watch:&lt;/strong&gt; The model names, participating agencies and operators, launch date, and contract terms all remain unconfirmed&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Google Cloud CEO visits Japan as AI cyber defense expands
&lt;/h2&gt;

&lt;p&gt;Google will give the Japanese government and major critical infrastructure operators early access to its latest AI models, &lt;a href="https://asia.nikkei.com/spotlight/cybersecurity/google-to-give-japan-early-access-to-frontier-ai-for-cybersecurity" rel="noopener noreferrer"&gt;Nikkei Asia reported&lt;/a&gt;. The aim is to strengthen Japan's defenses as the country faces a sharp rise in cyberattacks.&lt;/p&gt;

&lt;p&gt;Google Cloud CEO &lt;strong&gt;Thomas Kurian visited Japan on November 16&lt;/strong&gt; and discussed cybersecurity strategy with Japanese government officials. The move appears to expand to Japan the limited-access Fairwind Program that Google &lt;a href="https://joho-todai.com/google-gemini-3-8-flash-update/" rel="noopener noreferrer"&gt;announced in September (in Japanese)&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;When Google &lt;a href="https://blog.google/intl/ja-jp/company-news/technology/gemini-4-carbon/" rel="noopener noreferrer"&gt;announced&lt;/a&gt; the general availability of Gemini 4 Carbon on November 11, it said that selected defenders had received &lt;strong&gt;early access through Fairwind two weeks earlier&lt;/strong&gt;. It has not been confirmed whether Carbon is the model that will be offered in Japan.&lt;/p&gt;

&lt;h2&gt;
  
  
  How Fairwind works: Giving defenders a head start
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://deepmind.google/fairwind-program/" rel="noopener noreferrer"&gt;Fairwind Program&lt;/a&gt; was announced on September 2, 2026, as a program that gives defenders at governments, healthcare providers, telecommunications companies, and other critical infrastructure operators early access to Google's latest AI models. It is designed to let defenders act before new threats spread.&lt;/p&gt;

&lt;p&gt;Some partners receive exclusive access to the cybersecurity-focused "Gemini for cyber" model. Combined with CodeMender, the model can discover software vulnerabilities and automate fixes. CodeMender is a specialized code security agent that autonomously inspects and repairs code, reducing the burden of building and maintaining custom testing environments.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://joho-todai.com/gemini-4-carbon-ai-benchmark-top-ranking/" rel="noopener noreferrer"&gt;Gemini 4 Carbon for cyber (in Japanese)&lt;/a&gt; ranked &lt;strong&gt;first in all three benchmarks&lt;/strong&gt; for vulnerability discovery (SEC-bench Pro), exploit validation (ExploitGym), and patching (CWE-bench v1). Its ability to handle discovery, validation, and remediation in a single workflow is a defining feature.&lt;/p&gt;

&lt;p&gt;Fairwind imposes strict conditions. Use is limited to defensive and academic research, while reverse engineering and malware analysis are permitted only for defensive purposes. Sharing, redistribution, and resale of access are prohibited. Participating organizations must use multifactor authentication and access controls, and only their internal cybersecurity, incident response, and penetration-testing teams may access the models. Google says it works with &lt;strong&gt;thousands of partners&lt;/strong&gt; globally.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/google-fairwind-ai-japan-cybersecurity/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Japan's backdrop: 10 days after ¥10 billion in emergency measures
&lt;/h2&gt;

&lt;p&gt;On November 6, the Japanese government adopted an &lt;a href="https://joho-todai.com/government-cyber-emergency-100-billion-yen-ai-vulnerability/" rel="noopener noreferrer"&gt;emergency cybersecurity package (in Japanese)&lt;/a&gt; and allocated &lt;strong&gt;about ¥10 billion in reserve funds&lt;/strong&gt;. Prime Minister Sanae Takaichi set out four pillars: comprehensive basic measures using AI, stronger intelligence-gathering, analysis, and investigative capabilities, a strategic response to cyberattacks, and support for victims and prevention of further damage.&lt;/p&gt;

&lt;p&gt;Takaichi called for reports from critical infrastructure operators to be shared &lt;strong&gt;"within a few hours if possible, and at least on the same day."&lt;/strong&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Please aim to ensure that information that a critical infrastructure operator has been attacked is communicated within a few hours if possible, and at least on the same day.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This is from &lt;a href="https://www.kantei.go.jp/jp/105/actions/202611/06kaigi_cyber.html" rel="noopener noreferrer"&gt;remarks&lt;/a&gt; published by the Prime Minister's Office on November 6. The timeframe was a target set by the prime minister for consolidating information; it did not change the statutory reporting deadline.&lt;/p&gt;

&lt;p&gt;The emergency measures also call for earlier AI-assisted vulnerability assessments and stronger collection of attack intelligence through cooperation with overseas technology companies. Expanding Fairwind could become one concrete form of that cooperation.&lt;/p&gt;

&lt;h2&gt;
  
  
  What remains unknown
&lt;/h2&gt;

&lt;p&gt;Nikkei Asia reported that access would expand to Japan, but the specific terms have not been finalized. The names of the AI models, participating government agencies and operators, launch date, and contract terms all remain unconfirmed.&lt;/p&gt;

&lt;p&gt;Fairwind's general framework can be verified from public information. But announcements are still needed to establish the scope and terms under which Japanese government agencies and critical infrastructure operators will actually participate.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/google-fairwind-ai-japan-cybersecurity/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/google-fairwind-ai-japan-cybersecurity/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>ai</category>
    </item>
    <item>
      <title>Microsoft 365 Outage Leaves 15 With Chlorine Gas Poisoning at German Hotel After Alert Email Fails</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Mon, 16 Nov 2026 15:00:36 +0000</pubDate>
      <link>https://dev.to/joho_todai/microsoft-365-outage-leaves-15-with-chlorine-gas-poisoning-at-german-hotel-after-alert-email-fails-2g32</link>
      <guid>https://dev.to/joho_todai/microsoft-365-outage-leaves-15-with-chlorine-gas-poisoning-at-german-hotel-after-alert-email-fails-2g32</guid>
      <description>&lt;p&gt;An Exchange Online outage prevented a chlorine-level alert email from reaching staff at a German hotel during the Microsoft 365 disruption. Fifteen guests who used the pool were affected; WDR reported that 13 experienced respiratory irritation and four were taken to a hospital.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; A pool chlorine-dosing system malfunctioned at a Phantasialand hotel, and an alert email failed to arrive during the Microsoft 365 outage, leaving 15 people with chlorine gas poisoning&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why the harm escalated:&lt;/strong&gt; The safety system warning of abnormal chlorine levels depended solely on email notifications, and the Exchange Online outage completely severed the alert path&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Design question:&lt;/strong&gt; A cloud outage that halts work and one that causes people to collapse are fundamentally different. If safety alerts rely on an external service, the design must include a fallback for an outage&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The alert that never arrived
&lt;/h2&gt;

&lt;p&gt;At Phantasialand, a theme park in the western German city of Brühl, chlorine levels rose abnormally in the pool at Hotel Ling Bao on the afternoon of Friday, November 13 CET.&lt;/p&gt;

&lt;p&gt;The pool is equipped with a system that automatically adds chlorine. According to &lt;a href="https://www1.wdr.de/nrw/rheinland/rhein-erft-kreis/technischer-defekt-chlor-phantasialand-100.html" rel="noopener noreferrer"&gt;WDR, which interviewed multiple official sources&lt;/a&gt;, the automatic dosing system malfunctioned and &lt;strong&gt;pumped excessive chlorine into the pool&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The sauna area's "maintenance safety system" (Wartungssicherheitssystem) detected that the chlorine level had exceeded the threshold and attempted to send an alert email to staff. But the major Microsoft 365 outage was still underway. Exchange Online was down, and the email never arrived.&lt;/p&gt;

&lt;p&gt;Phantasialand told WDR the following day, November 14, that the outage had prevented delivery of the email. Guests continued using the pool while no one was aware of the abnormal chlorine level.&lt;/p&gt;

&lt;p&gt;The guests themselves were the first to report that something was wrong. After using the pool, a succession of guests complained of respiratory irritation and other health problems. Fire and emergency services were called, and 40 emergency responders were dispatched to the scene.&lt;/p&gt;

&lt;p&gt;Tests by emergency personnel confirmed a defect in the automatic dosing system and &lt;strong&gt;significantly elevated chlorine levels&lt;/strong&gt;. Ultimately, 15 people were triaged and four were taken to a hospital. One person had moderate injuries and remained hospitalized for observation. The pool was temporarily closed and reopened after all the water was replaced.&lt;/p&gt;

&lt;h2&gt;
  
  
  Safety built on email
&lt;/h2&gt;

&lt;p&gt;The core issue in this incident is that a safety-related monitoring function depended on email as its communications path.&lt;/p&gt;

&lt;p&gt;Monitoring chlorine levels in a pool is a safety function directly connected to human life. High chlorine concentrations can cause respiratory injury. Fifteen people were affected in this incident, 13 of whom reported respiratory irritation. The alert was carried over a cloud email service via the internet.&lt;/p&gt;

&lt;p&gt;Email is not a communications method that is "guaranteed to arrive." A sending-server failure, a receiving-server failure, a network interruption, a false positive from a spam filter, or a recipient who does not read the message can make an alert disappear at any stage. This time, one of those failure points—a worldwide Exchange Online outage—occurred in reality.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://joho-todai.com/microsoft-365-outlook-teams-sharepoint-outage/" rel="noopener noreferrer"&gt;Microsoft 365 outage began at 2:05 a.m. PST on November 12 (in Japanese)&lt;/a&gt; and &lt;a href="https://joho-todai.com/microsoft-365-outage-personal-email-fixed/" rel="noopener noreferrer"&gt;recovery was declared on November 14 (in Japanese)&lt;/a&gt;. It was caused by &lt;a href="https://joho-todai.com/microsoft-365-outage-network-upgrade-failure/" rel="noopener noreferrer"&gt;a defect introduced during an earlier network-capacity expansion (in Japanese)&lt;/a&gt; that surfaced when another change was made. During the outage, Deutsche Bahn was unable to send timetables for diverted services, while some workplaces in Japan reverted to fax.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/microsoft-365-outage-hotel-chlorine-gas-poisoning/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Email going down and delaying work is fundamentally different from email going down and &lt;strong&gt;causing people to inhale chlorine gas&lt;/strong&gt;. The former is an inconvenience; the latter causes injury. This incident showed that where an outage spreads is determined not by the outage itself, but by the design of the systems that depend on it.&lt;/p&gt;

&lt;p&gt;If safety-related alerts depend on an external cloud service, the design process must ask what happens when that service stops. A locally sounding alarm, a physical warning device, and regular human inspections all already exist as technologies. The question is whether the design decision was made to include them.&lt;/p&gt;

&lt;p&gt;At the Phantasialand hotel, the pool water was replaced and operations resumed. Of the 15 people assessed by emergency services, 14 had minor injuries and one was classified as moderately injured. Four were taken to a hospital, and one remained hospitalized for observation. But replacing the water does not erase the fact that people collapsed because an email did not arrive.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/microsoft-365-outage-hotel-chlorine-gas-poisoning/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/microsoft-365-outage-hotel-chlorine-gas-poisoning/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>microsoft</category>
    </item>
    <item>
      <title>CachyOS Revamps Its Package Management UI Again, Unveils Package Installer 2.0</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Sun, 15 Nov 2026 15:00:21 +0000</pubDate>
      <link>https://dev.to/joho_todai/cachyos-revamps-its-package-management-ui-again-unveils-package-installer-20-30of</link>
      <guid>https://dev.to/joho_todai/cachyos-revamps-its-package-management-ui-again-unveils-package-installer-20-30of</guid>
      <description>&lt;p&gt;CachyOS has previewed Package Installer 2.0, which manages AUR, Flatpak, and official repositories in one interface and supports nine color themes. It is the distribution's third package management UI overhaul of 2026.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; CachyOS previewed Package Installer 2.0, which manages AUR, Flatpak, and official repositories in one interface and supports nine color themes&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; This is the third package management UI overhaul of 2026, following Octopi, Shelly, and now Package Installer 2.0. CachyOS is repeatedly lowering the barriers for newcomers to Arch-based Linux&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to watch:&lt;/strong&gt; The relationship between Shelly and Package Installer 2.0 remains unclear, as does whether version 2.0 will be included in the new CachyOS release teased for the near future&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  A third rebrand in 2026
&lt;/h2&gt;

&lt;p&gt;CachyOS &lt;a href="https://x.com/CachyOS/status/2121300731315507696" rel="noopener noreferrer"&gt;published&lt;/a&gt; an early preview of CachyOS Package Installer 2.0 on November 14. The Arch Linux-based distribution is popular with gamers and &lt;a href="https://joho-todai.com/cachyos-surpasses-steamos-linux-september/" rel="noopener noreferrer"&gt;surpassed SteamOS in Linux share for the first time (in Japanese)&lt;/a&gt; in the September Steam Hardware Survey. But its Arch foundation itself remains a barrier for first-time Linux users. Package installation epitomizes that difficulty, and CachyOS has &lt;strong&gt;worked repeatedly throughout 2026 to lower this barrier&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;In April, it &lt;a href="https://joho-todai.com/cachyos-april-shelly-vram-management-improvements/" rel="noopener noreferrer"&gt;replaced (in Japanese)&lt;/a&gt; the long-standing, text-oriented Octopi GUI package manager with Shelly. Shelly was a GTK 4-based tool—a Linux UI development framework—that could manage AUR, Flatpak, and AppImage packages together, and it was written in C# and .NET 10. By August, however, CachyOS was &lt;a href="https://joho-todai.com/cachyos-shelly-rewritten-zig-server-edition/" rel="noopener noreferrer"&gt;rewriting Shelly entirely in Zig (in Japanese)&lt;/a&gt; to eliminate the managed runtime and move to a native binary.&lt;/p&gt;

&lt;p&gt;Now comes Package Installer 2.0, only three months after Shelly was rewritten in Zig. CachyOS Package Installer already had a Qt6-based v1.x series, and version 2.0 is its major update.&lt;/p&gt;

&lt;p&gt;CachyOS has not disclosed how Shelly and Package Installer 2.0 will relate to each other—whether they will coexist or one will become the default. Users have also asked what happened to Shelly and whether it was not supposed to become the primary package installer.&lt;/p&gt;

&lt;p&gt;The project also teased that &lt;strong&gt;a new CachyOS release is coming soon&lt;/strong&gt;. Whether Package Installer 2.0 will be included remains undecided.&lt;/p&gt;

&lt;h2&gt;
  
  
  What will change
&lt;/h2&gt;

&lt;p&gt;Package Installer 2.0 is being redesigned around an app-store layout similar to Bazaar, the standard app store in the gaming-focused Linux distribution Bazzite and a front end for the Flathub Linux application distribution platform. Instead of the tabular list used by conventional package managers, the new interface is expected to let users install applications while viewing preview images and descriptions.&lt;/p&gt;

&lt;p&gt;Supported package sources include the official CachyOS repositories, &lt;strong&gt;AUR&lt;/strong&gt; (the Arch User Repository, where Arch users share packages), and &lt;strong&gt;Flatpak&lt;/strong&gt; (a sandboxed package format that works across distributions).&lt;/p&gt;

&lt;p&gt;They can be managed together in one interface. Flatpak and AUR both use opt-in toggles and are disabled by default. AppImage support is also planned, though as an installer rather than as a managed catalog.&lt;/p&gt;

&lt;p&gt;There are nine color themes: four CachyOS themes plus Nord, Dracula, Catppuccin, Tokyo Night, and Solarized. Bazaar does not have a theme feature, giving Package Installer 2.0 a point of differentiation.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/cachyos-package-installer-2-0-redesign/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The development team also addressed its direction in replies. Features will be added over time. The team wants the interface to be as user-friendly as possible but does not intend to overwhelm it with information.&lt;/p&gt;

&lt;p&gt;Arch-based distributions traditionally make users choose performance at the cost of accessibility. CachyOS is breaking down those barriers one by one while preserving its performance advantages. Revising a package management tool three times in seven months is fast, but that speed is itself part of CachyOS's character.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/cachyos-package-installer-2-0-redesign/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/cachyos-package-installer-2-0-redesign/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>linux</category>
    </item>
    <item>
      <title>AI-Powered Penetration Testing Traffic Surges More Than 30x, Cloudflare Says, as Stolen Tokens Enable Internal Breaches</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Sat, 14 Nov 2026 15:00:33 +0000</pubDate>
      <link>https://dev.to/joho_todai/ai-powered-penetration-testing-traffic-surges-more-than-30x-cloudflare-says-as-stolen-tokens-4mld</link>
      <guid>https://dev.to/joho_todai/ai-powered-penetration-testing-traffic-surges-more-than-30x-cloudflare-says-as-stolen-tokens-4mld</guid>
      <description>&lt;p&gt;Cloudflare reported that traffic from AI penetration-testing tools has grown more than 30-fold since October 1. More than 12,000 ARTEX-related repositories now exist, and real-world breaches have been confirmed when the tools are combined with stolen CI/CD tokens.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Cloudflare reported that requests from AI penetration-testing tools increased more than 30-fold between October 1 and November 12, with more than 50 tools and thousands of IP addresses observed&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; Stolen CI/CD tokens were used in combination with ARTEX, and Cloudflare confirmed an incident in which numerous Worker scripts were created inside a Cloudflare environment through an API token without IP restrictions&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to do:&lt;/strong&gt; External defenses such as a WAF are not enough; organizations must reduce API-token privileges, set expiration dates, and separate credentials by environment&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  More than 30x the requests and over 50 tools
&lt;/h2&gt;

&lt;p&gt;Cloudforce One, Cloudflare's threat research team, &lt;a href="https://blog.cloudflare.com/adapting-to-ai-penetration-testing-orchestrators/" rel="noopener noreferrer"&gt;published a report&lt;/a&gt; on AI-driven penetration-testing tools on November 14 Japan time.&lt;/p&gt;

&lt;p&gt;For a representative group of AI penetration-testing tools tracked by the team, &lt;strong&gt;daily request volume grew by more than 30x between October 1 and November 12&lt;/strong&gt;. In the latest three-day period alone, more than 50 frameworks attempted to scan Cloudflare customers, originating from thousands of IP addresses.&lt;/p&gt;

&lt;p&gt;However, the 30-fold figure does not mean that the total number of cyberattacks worldwide increased by 30 times. It is the growth rate in requests from a specific group of AI penetration-testing tools that Cloudflare tracks on its own network, and multiple tools may be used from the same IP address. The number of IP addresses, types of tools, scanning attempts, and actual breaches must be treated as separate metrics.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/ai-penetration-testing-traffic-spike-cloudflare/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;At the center of the surge is the AI penetration-testing tool ARTEX and its derivatives, which Joho no Todai has &lt;a href="https://joho-todai.com/storm-3358-exploits-southeast-asia-banks-artex-ai/" rel="noopener noreferrer"&gt;covered repeatedly in earlier reporting (in Japanese)&lt;/a&gt;. As of November 12, ARTEX and its forks accounted for &lt;strong&gt;more than 60%&lt;/strong&gt; of the tools Cloudflare was tracking.&lt;/p&gt;

&lt;h2&gt;
  
  
  Deletion fueled proliferation
&lt;/h2&gt;

&lt;p&gt;ARTEX developer Autumn-27 deleted the public repository on October 8, saying that its misuse ran counter to the original purpose.&lt;/p&gt;

&lt;p&gt;But repositories containing copied code and independently modified versions did not disappear. Cloudflare said there are &lt;strong&gt;more than 12,000 related repositories&lt;/strong&gt;, including hundreds of public derivatives with their own code changes. In a lineage map of derivatives analyzed by Cloudflare, 204 public repositories contained independent code changes.&lt;/p&gt;

&lt;p&gt;Cloudflare calls the phenomenon the hydra effect. Deleting the original repository, combined with widespread reporting of the deletion, instead drove a surge in forks and modified versions. In about five weeks after the deletion, requests from AI penetration-testing tools including ARTEX grew more than tenfold and then surpassed 30 times their previous level.&lt;/p&gt;

&lt;p&gt;ARTEX-style tools use large language models (LLMs) to divide work among multiple AI agents. They continuously gather information about targets, enumerate services, check for vulnerabilities, and choose what to investigate next. Because Model Context Protocol (MCP) can connect existing security-testing tools, the AI does not need to write original attack code each time. It can combine existing tools for a given purpose and use information gathered along the way to make its next decision.&lt;/p&gt;

&lt;p&gt;Representative tools cited by Cloudflare include LuaN1aoAgent, Decepticon, Cairn, and PentestGPT-Agent in addition to ARTEX. They share an approach in which multiple AI agents divide roles and manage goals and findings through graphs or task lists. Compared with a scanner that checks only a single vulnerability, they can more easily expand an investigation from one discovered weakness to another system or action.&lt;/p&gt;

&lt;h2&gt;
  
  
  Stolen tokens and ARTEX combined in a breach
&lt;/h2&gt;

&lt;p&gt;The most important part of the report is a case Cloudforce One confirmed during a real incident response.&lt;/p&gt;

&lt;p&gt;In multiple incidents handled by the team over the past eight weeks, &lt;strong&gt;non-expiring static access tokens used in CI/CD pipelines were compromised&lt;/strong&gt;, giving attackers broad privileges across multiple cloud environments.&lt;/p&gt;

&lt;p&gt;In one specific case, ARTEX used a compromised account-level API token originating from a third-party hyperscale cloud environment. The token had neither IP-address restrictions nor request limits.&lt;/p&gt;

&lt;p&gt;The attacker created numerous new Worker scripts, programs that run in Cloudflare's serverless environment. They were given role-specific names including artex-&lt;em&gt;, artex-pgrecon-&lt;/em&gt;, artex-pgw-*, and artex-do-*, with multiple functions operated in parallel.&lt;/p&gt;

&lt;p&gt;Observed activity included more than 69 requests per minute and 400 database-dump requests in one hour. Cloudflare also observed attempts to establish persistence and verify whether session tokens could be used with other services. The figure of 400 refers to the number of dump requests, not to the successful acquisition of data from 400 databases.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/ai-penetration-testing-traffic-spike-cloudflare/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Cloudforce One said much of the activity may have been guided to some extent by a human operator.&lt;/p&gt;

&lt;p&gt;The attack was possible not only because of the AI agents' capabilities, but also because the compromised token remained valid for a long time, had broad privileges, and lacked source-IP restrictions. A WAF (web application firewall) can block external vulnerability probing, but API operations performed with legitimate credentials take place outside that boundary.&lt;/p&gt;

&lt;p&gt;When CI/CD tokens have broad privileges across multiple production services, a development-environment compromise can become a path into production systems. Tools such as ARTEX can test this path faster than a person and across more services.&lt;/p&gt;

&lt;h2&gt;
  
  
  Protect both the perimeter and the inside
&lt;/h2&gt;

&lt;p&gt;Cloudflare proposed six areas of defense: WAF and rate limiting, anomaly detection, bot protection, authenticated access controls, API schema validation, and threat intelligence. The company calls the recommended architecture its "Modern Security Model," organizing capabilities that were already available.&lt;/p&gt;

&lt;p&gt;These areas correspond to Cloudflare's own products and therefore carry a commercial context. The underlying ideas, however, remain useful beyond the company's product portfolio.&lt;/p&gt;

&lt;p&gt;For administrators of public systems, the practical priority is to design defenses on the assumption that attackers will combine multiple paths. Specifically, this means:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Rate-limit large volumes of probing requests&lt;/li&gt;
&lt;li&gt;Validate input to reject unexpected API calls&lt;/li&gt;
&lt;li&gt;Require authentication for administrative interfaces&lt;/li&gt;
&lt;li&gt;Reduce API-token privileges to the minimum required&lt;/li&gt;
&lt;li&gt;Set token expiration dates and reissue tokens regularly&lt;/li&gt;
&lt;li&gt;Restrict the source IP addresses allowed to use tokens&lt;/li&gt;
&lt;li&gt;Separate credentials across development, testing, and production environments&lt;/li&gt;
&lt;li&gt;Maintain the ability to detect anomalous privilege use in access logs&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The case shows that even if a WAF reduces external probing, a stolen privileged token can provide another route into the environment. Neither perimeter defenses nor designs that limit post-compromise privilege expansion are sufficient on their own.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI changed the "speed and breadth"
&lt;/h2&gt;

&lt;p&gt;AI-powered penetration testing is also used for legitimate security assessments that organizations authorize themselves. Not every automated scan observed by Cloudflare was necessarily illegal intrusion activity, and the use of a particular tool or IP address alone cannot establish malicious intent.&lt;/p&gt;

&lt;p&gt;The change revealed here is not that AI has begun inventing unknown vulnerabilities automatically. It is that open-source toolkits are spreading the ability to use existing testing tools continuously, choose the next action based on findings, and examine multiple environments in a short time.&lt;/p&gt;

&lt;p&gt;The harmful use of ARTEX has been confirmed across multiple environments over the past two months, including &lt;a href="https://joho-todai.com/korean-banks-ai-attack-authentication-vulnerability/" rel="noopener noreferrer"&gt;South Korean financial institutions (in Japanese)&lt;/a&gt;, &lt;a href="https://joho-todai.com/connectwise-ai-agents-microsoft-365-auth-attacks/" rel="noopener noreferrer"&gt;authentication attacks against Microsoft 365 tenants (in Japanese)&lt;/a&gt;, &lt;a href="https://joho-todai.com/github-leaked-credentials-azure-breach-5-minutes/" rel="noopener noreferrer"&gt;an Azure Key Vault compromise (in Japanese)&lt;/a&gt;, and &lt;a href="https://joho-todai.com/storm-3358-exploits-southeast-asia-banks-artex-ai/" rel="noopener noreferrer"&gt;financial institutions in Southeast Asia (in Japanese)&lt;/a&gt;. Cloudflare's latest report adds quantitative evidence of the traffic surge and a new path involving chained credentials.&lt;/p&gt;

&lt;p&gt;When the repository disappeared on October 8, ARTEX was supposed to be finished. The reality that 12,000 copies remain and more than 50 successors and modified versions have begun running demonstrates in numbers that "deletion" is not an effective remedy for open-source penetration-testing tools.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/ai-penetration-testing-traffic-spike-cloudflare/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/ai-penetration-testing-traffic-spike-cloudflare/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>ai</category>
    </item>
    <item>
      <title>Rust 1.100 Released: The Never Type Stabilizes After a Decade, While Cargo Adds Supply Chain Protection</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Fri, 13 Nov 2026 15:00:05 +0000</pubDate>
      <link>https://dev.to/joho_todai/rust-1100-released-the-never-type-stabilizes-after-a-decade-while-cargo-adds-supply-chain-5e90</link>
      <guid>https://dev.to/joho_todai/rust-1100-released-the-never-type-stabilizes-after-a-decade-while-cargo-adds-supply-chain-5e90</guid>
      <description>&lt;p&gt;Rust 1.100.0 was released on November 12. The release marks major milestones across the language and toolchain, led by stabilization of the never type after five failed attempts over 10 years, the foundation of the Allocator API, and a dependency cooldown feature in Cargo.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Rust 1.100 was released on November 12, stabilizing the never type (!) across all editions after five failed attempts over 10 years&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it was difficult:&lt;/strong&gt; The plan to make Infallible an alias for the never type required a breaking change to fallback rules, and pre-release testing found 3,277 regressions&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Practical progress:&lt;/strong&gt; Alongside stabilization of the Allocator API foundation, Cargo gained the min-publish-age feature, which automatically holds newly published packages and adds another supply chain defense option&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  10 years, five failed attempts, and the never type
&lt;/h2&gt;

&lt;p&gt;Rust has a special type written as &lt;code&gt;!&lt;/code&gt;. Known as the &lt;strong&gt;never type&lt;/strong&gt;, it represents in the type system that a value can never exist. Expressions that do not return, such as &lt;code&gt;return&lt;/code&gt;, &lt;code&gt;break&lt;/code&gt;, and &lt;code&gt;std::process::exit&lt;/code&gt;, all have this &lt;code&gt;!&lt;/code&gt; type. Because no value can exist, &lt;code&gt;!&lt;/code&gt; can be converted to any type. This property allows type checking to succeed even when one arm of a &lt;code&gt;match&lt;/code&gt; ends with &lt;code&gt;return&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The concept of the never type had long existed internally in Rust. A function return type could be written as &lt;code&gt;-&amp;gt; !&lt;/code&gt; to declare that the function never returns to its caller. However, it could not be used freely as a type parameter—for example, as &lt;code&gt;Result&amp;lt;T, !&amp;gt;&lt;/code&gt; to express that a Result can never fail.&lt;/p&gt;

&lt;p&gt;Rust 1.100 removes this restriction. &lt;code&gt;!&lt;/code&gt; can now be &lt;a href="https://blog.rust-lang.org/2026/11/12/Rust-1.100.0/" rel="noopener noreferrer"&gt;written in any type position&lt;/a&gt;, completing the never-type project goal.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why did it take 10 years?
&lt;/h3&gt;

&lt;p&gt;The tracking issue for the original proposal, RFC 1216 (#35121), was opened in July 2016. Since then, &lt;strong&gt;all five stabilization attempts&lt;/strong&gt; have failed.&lt;/p&gt;

&lt;p&gt;The core difficulty was integration with &lt;code&gt;Infallible&lt;/code&gt;. Rust's standard library includes a type called &lt;code&gt;std::convert::Infallible&lt;/code&gt; that represents an operation that cannot fail. &lt;code&gt;Infallible&lt;/code&gt; was introduced with the plan to eventually make it an alias for the never type, but doing so required changing the fallback rules used when the compiler cannot infer a type.&lt;/p&gt;

&lt;p&gt;Previously, the compiler automatically assigned the empty tuple &lt;code&gt;()&lt;/code&gt; where it could not determine a type. This had to be changed to &lt;code&gt;!&lt;/code&gt; to unify Infallible and the never type. However, changing this fallback can alter the behavior of existing code, making it a breaking change.&lt;/p&gt;

&lt;p&gt;Rust adopted a phased strategy: switch the fallback to &lt;code&gt;!&lt;/code&gt; first in the Rust 2024 edition, allow the change to spread through the ecosystem, and then apply it to all editions—2015, 2018, and 2021—in version 1.100. Pre-release testing against crates published on crates.io initially found 3,277 regressions. After fixes spanning 18 commits, compatibility patches were also distributed in advance to affected public crates.&lt;/p&gt;

&lt;p&gt;If updating a dependency breaks code, an update to that crate may already provide a fix. The compiler may also mark a broader range of code as unreachable.&lt;/p&gt;

&lt;h2&gt;
  
  
  Allocator API: More options for memory management
&lt;/h2&gt;

&lt;p&gt;Another major stabilization is the foundation of the Allocator API. Long restricted to nightly under &lt;code&gt;#![feature(allocator_api)]&lt;/code&gt;, the &lt;code&gt;Allocator&lt;/code&gt; trait has entered stable Rust, allowing library authors to define custom memory allocators and pass them to &lt;code&gt;Vec&lt;/code&gt; and &lt;code&gt;Box&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Many systems, including game engines, databases, and embedded systems, need fine-grained control over memory allocation. Until now, developers had to replace the global allocator for the entire system or manage memory themselves with unsafe code. The Allocator API allows different allocators to be assigned safely on a per-type basis.&lt;/p&gt;

&lt;p&gt;Only &lt;code&gt;Vec&lt;/code&gt; and &lt;code&gt;Box&lt;/code&gt; are currently supported. Expansion to other collection types, including &lt;code&gt;HashMap&lt;/code&gt; and &lt;code&gt;Rc&lt;/code&gt;, is planned as future work under the Allocators 1.0 project goal.&lt;/p&gt;

&lt;h2&gt;
  
  
  DropGuard and conjure_zst: Two small but practical new APIs
&lt;/h2&gt;

&lt;h3&gt;
  
  
  DropGuard
&lt;/h3&gt;

&lt;p&gt;&lt;code&gt;core::mem::DropGuard&lt;/code&gt; has been added as a new helper type that guarantees a closure, or function object, will run when a scope is exited. It accepts a variable and a closure and implements &lt;code&gt;Deref&lt;/code&gt; and &lt;code&gt;DerefMut&lt;/code&gt;, so the wrapped variable can continue to be used directly.&lt;/p&gt;

&lt;p&gt;A typical use is implementing transactional rollback: if an operation fails partway through, the DropGuard closure automatically reverts the state. If it succeeds, the &lt;code&gt;dismiss&lt;/code&gt; method cancels execution of the closure and commits the change.&lt;/p&gt;

&lt;p&gt;Rust's &lt;code&gt;Drop&lt;/code&gt; trait can accomplish the same result, but the practical benefit is that DropGuard needs only a closure and does not require defining a type.&lt;/p&gt;

&lt;h3&gt;
  
  
  conjure_zst
&lt;/h3&gt;

&lt;p&gt;&lt;code&gt;std::mem::conjure_zst&lt;/code&gt; has been added as an official &lt;code&gt;unsafe&lt;/code&gt; function for creating values of zero-sized types (ZSTs). A type with no size has no representation in memory, and only one value can exist. Previously, developers created such values through techniques such as &lt;code&gt;transmute(())&lt;/code&gt; or &lt;code&gt;MaybeUninit::uninit().assume_init()&lt;/code&gt;, but these were workarounds outside the intended purpose of those APIs.&lt;/p&gt;

&lt;p&gt;The function is &lt;code&gt;unsafe&lt;/code&gt; because some APIs ensure safety by intentionally preventing construction of a zero-sized type; such types must not be created without a valid reason. The official blog gives PyO3's &lt;code&gt;Python&amp;lt;'py&amp;gt;&lt;/code&gt; token as an example. This type is a marker indicating that Python's Global Interpreter Lock (GIL) is held. If the token could be freely created, Python APIs could be called without holding the GIL.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three Cargo improvements
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Detecting unused dependencies
&lt;/h3&gt;

&lt;p&gt;Cargo now reports a warning for dependencies listed in &lt;code&gt;Cargo.toml&lt;/code&gt; but unused by every build target. It was introduced as part of the new &lt;code&gt;[lints.cargo]&lt;/code&gt; lint system, completing this project goal as well.&lt;/p&gt;

&lt;p&gt;There is one limitation: the warning appears only when all build targets in the package are built. If only a specific binary is built, as with &lt;code&gt;cargo check --bin foo&lt;/code&gt;, Cargo does not warn about dependencies used by other binaries. Unused &lt;code&gt;dev-dependencies&lt;/code&gt; are also excluded from detection.&lt;/p&gt;

&lt;h3&gt;
  
  
  Reducing cache collisions
&lt;/h3&gt;

&lt;p&gt;Cargo's build-directory layout has been redesigned so that the build-configuration hash is applied consistently to every artifact. Platform constraints previously meant that some artifacts received no hash, causing unnecessary rebuilds, missed rebuilds, and confusion when building across git worktrees. The work began with the 2025 build directory layout project goal and opens the way for future cache garbage collection and sharing across workspaces.&lt;/p&gt;

&lt;p&gt;The change may affect tools and tests that depend on the internal structure of the build directory.&lt;/p&gt;

&lt;h3&gt;
  
  
  min-publish-age: A dependency cooldown
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://joho-todai.com/rust-developer-targeted-attack-arrayref-incident/" rel="noopener noreferrer"&gt;Malicious package versions (in Japanese)&lt;/a&gt; are often discovered by automated scanners or the community within hours or days of publication. &lt;strong&gt;The min-publish-age feature&lt;/strong&gt; prevents new versions from being pulled in automatically until that detection window has passed.&lt;/p&gt;

&lt;p&gt;It can be configured in &lt;code&gt;.cargo/config.toml&lt;/code&gt; as follows.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight toml"&gt;&lt;code&gt;&lt;span class="k"&gt;[registry]&lt;/span&gt;
&lt;span class="na"&gt;global-min-publish-age&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"3 days"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When &lt;code&gt;cargo update&lt;/code&gt; runs, versions newer than the configured period are automatically held back, and Cargo displays what was withheld. If a new version is needed urgently, the restriction can be overridden for that run with the &lt;code&gt;CARGO_RESOLVER_INCOMPATIBLE_PUBLISH_AGE=allow&lt;/code&gt; environment variable.&lt;/p&gt;

&lt;p&gt;The restriction does not apply to &lt;code&gt;cargo install&lt;/code&gt;. Versions without a publication-time field in the registry index are also excluded.&lt;/p&gt;

&lt;p&gt;The official blog says the feature reduces the risk of supply chain attacks but is not a complete defense on its own. Quarantine procedures for crates and automatic detection of suspicious releases are also planned under the crates.io supply chain hardening project goal.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/rust-1-100-never-type-stable-cargo-supply-chain/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  What the 101st stable release demonstrates
&lt;/h2&gt;

&lt;p&gt;Rust 1.100 may look like a milestone by version number, but the substance marks a deeper transition. Stabilizing the never type after a decade demonstrates Rust's stabilization method itself: not avoiding changes that might break code, but measuring the scope of breakage, fixing it, and migrating in stages. Stabilizing the Allocator API increases freedom in memory management, while Cargo's min-publish-age draws a defensive line across the vulnerable period in the software supply chain.&lt;/p&gt;

&lt;p&gt;Language maturity and practical safety have arrived in the same release.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/rust-1-100-never-type-stable-cargo-supply-chain/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/rust-1-100-never-type-stable-cargo-supply-chain/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
    </item>
    <item>
      <title>FreeBSD Surveys 140 Developers on AI: 16% Favor a Ban, 91% Prioritize Human Accountability</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Thu, 12 Nov 2026 15:00:28 +0000</pubDate>
      <link>https://dev.to/joho_todai/freebsd-surveys-140-developers-on-ai-16-favor-a-ban-91-prioritize-human-accountability-3h53</link>
      <guid>https://dev.to/joho_todai/freebsd-surveys-140-developers-on-ai-16-favor-a-ban-91-prioritize-human-accountability-3h53</guid>
      <description>&lt;p&gt;FreeBSD's Q3 2026 report includes an AI survey of 140 developers. Just 16% favor rejecting AI-assisted code outright. Meanwhile, AI scrapers keep attacking the project's infrastructure as developers work out the terms for accepting AI.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; In a survey of 140 FreeBSD developers, 88% wanted an AI policy and 16% favored outright rejection of AI-assisted code. Human accountability was the top principle, cited by 91%&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Where it matters in practice:&lt;/strong&gt; AI wrote the NFS over RDMA server code. Without a policy in place, it is available as an unofficial port&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI as an adversary:&lt;/strong&gt; Scrapers collecting LLM training data continually attack FreeBSD infrastructure through botnets. The operations team calls them "trainers of probabilistic malware"&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The author's view:&lt;/strong&gt; FreeBSD is debating when to accept AI in its code while defending infrastructure from AI-related attacks. The different roles AI plays inside one project reflect a wider tension in open source&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Developers are open to using AI
&lt;/h2&gt;

&lt;p&gt;FreeBSD Core Team members &lt;a href="https://www.freebsd.org/status/report-2026-07-2026-09/" rel="noopener noreferrer"&gt;surveyed project developers&lt;/a&gt; about their use of AI and large language models (LLMs). &lt;strong&gt;140 people&lt;/strong&gt; responded.&lt;/p&gt;

&lt;p&gt;The results show broad support for setting terms for AI use. &lt;strong&gt;88%&lt;/strong&gt; wanted some form of project AI policy. Of all respondents, 49% wanted a formal written policy and 39% thought lightweight guidelines would suffice. Only 6% wanted the project to explicitly discourage AI, and 3% said existing contribution rules were enough.&lt;/p&gt;

&lt;p&gt;Responses split four ways on how to handle code produced with AI assistance. The largest share, 35%, wanted mandatory disclosure. Another 25% wanted stricter review than for human-written code, and 22% favored treating it like any other contribution. The smallest group, &lt;strong&gt;16%&lt;/strong&gt;, wanted it rejected outright.&lt;/p&gt;

&lt;p&gt;In July 2026, GCC &lt;a href="https://joho-todai.com/gcc-llm-generated-code-guidelines/" rel="noopener noreferrer"&gt;formally adopted a policy (in Japanese)&lt;/a&gt; rejecting LLM-generated code longer than roughly 15 lines. In Void Linux, &lt;a href="https://joho-todai.com/void-linux-maintainer-llm-policy-violation/" rel="noopener noreferrer"&gt;a maintainer dropped 113 packages at once after being accused of violating its AI policy (in Japanese)&lt;/a&gt;. QEMU moved from a blanket ban to &lt;a href="https://joho-todai.com/qemu-ai-generated-code-ban-limited-permission/" rel="noopener noreferrer"&gt;allowing limited use under certain conditions (in Japanese)&lt;/a&gt;. As projects take different approaches, FreeBSD's survey shows what its developers themselves think.&lt;/p&gt;

&lt;h2&gt;
  
  
  Human accountability and transparency lead the way
&lt;/h2&gt;

&lt;p&gt;The breakdown shows that developers want conditions for using AI, with outright rejection attracting less support.&lt;/p&gt;

&lt;p&gt;The most popular safeguard was &lt;strong&gt;mandatory human review, at 78%&lt;/strong&gt;. Mandatory disclosure followed at 69%, documentation of provenance and licensing at 47%, and mandatory testing at 39%. Some 48% wanted AI use always disclosed and 36% wanted disclosure only for substantial involvement, giving &lt;strong&gt;84%&lt;/strong&gt; in favor of some disclosure. Marking AI use in commit logs drew 79% support, and naming the tool drew 56%.&lt;/p&gt;

&lt;p&gt;Asked which principles the project should prioritize, &lt;strong&gt;91%&lt;/strong&gt; chose human accountability, the top response. Transparency drew 78%. Maintaining code quality was the leading goal for the policy.&lt;/p&gt;

&lt;p&gt;Among quality concerns, correctness and subtle bugs came first at 39%. Copyright and licensing issues concerned 75% of respondents.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/freebsd-ai-usage-survey-developers/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;For AI-generated documentation, 65% would accept it with human review. Some 54% supported limited use of AI tools in project infrastructure, while 29% opposed it.&lt;/p&gt;

&lt;p&gt;The survey informed the creation of an AI working group with members holding a range of views. It held its kickoff meeting on October 7 and plans a final report in Q4.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI-written code already exists
&lt;/h2&gt;

&lt;p&gt;The policy debate is already affecting concrete work.&lt;/p&gt;

&lt;p&gt;The Q3 report's entry on NFS over RDMA (Remote Direct Memory Access) says generative AI produced the server-side code. Testing of the client side is going well, and it is expected to reach FreeBSD's main branch within weeks. The project cannot officially merge the server code while it lacks an AI policy, so the code is available as an unofficial port that builds a kernel module.&lt;/p&gt;

&lt;p&gt;The lack of a policy prevents an official merge, which creates a practical reason to write one. The code arrived before the rules for accepting it.&lt;/p&gt;

&lt;h2&gt;
  
  
  "Trainers of probabilistic malware"
&lt;/h2&gt;

&lt;p&gt;While FreeBSD debates how to accept AI contributions, scrapers collecting data for AI companies continue to attack its infrastructure.&lt;/p&gt;

&lt;p&gt;The Cluster Administration Team reports that scrapers collecting LLM training data are attacking FreeBSD.org infrastructure constantly. Scraping has shifted from company infrastructure to &lt;strong&gt;botnets&lt;/strong&gt; using residential proxies, which route traffic through household and business networks. This makes scraper traffic hard to distinguish from legitimate requests. The team says blocking it at the network edge risks substantial collateral damage.&lt;/p&gt;

&lt;p&gt;The main target is cgit.FreeBSD.org, the web interface for FreeBSD source code. The report calls the scrapers targeting it &lt;strong&gt;"trainers of probabilistic malware."&lt;/strong&gt; Traffic arrives continuously from compromised devices around the world.&lt;/p&gt;

&lt;p&gt;During Q3, the Cluster Administration Team deployed additional cgit mirrors in California and Paris to spread the load. It reported spending considerable time mitigating distributed denial-of-service (DDoS) attacks that quarter.&lt;/p&gt;

&lt;p&gt;The Bugmeister Team also lists this as its highest priority. It says the cause is known and the Core Team recognizes the severity, but Bugmeister has limited means to address it.&lt;/p&gt;

&lt;p&gt;A volunteer operations team defends a cluster with 143 physical machines and 324 jails, for 467 installations across 14 sites. The traffic comes at botnet scale. The resources on each side remain uneven.&lt;/p&gt;

&lt;h2&gt;
  
  
  EU compliance and software bills of materials
&lt;/h2&gt;

&lt;p&gt;Another source of pressure is the EU Cyber Resilience Act (CRA). Vulnerability and incident reporting requirements for manufacturers took effect on September 11, 2026. Reporting requirements for open-source software stewards begin on December 11, 2027.&lt;/p&gt;

&lt;p&gt;The FreeBSD Core Team published a statement recognizing the FreeBSD Foundation as the project's open-source software steward. The foundation has prepared a cybersecurity policy and CRA reporting protocol and published guidance for business users and community members on its website.&lt;/p&gt;

&lt;p&gt;A framework that automatically generates software bills of materials (SBOMs) at build time has entered review. An amd64 build produces &lt;strong&gt;218 artifacts&lt;/strong&gt;, with one for each base package and each library actually used in the build, enabling the full dependency chain to be traced. It supports SPDX 2.2 and Lite 3.0.1 output, with CycloneDX support underway. SBOM support for the ports tree is undergoing separate review.&lt;/p&gt;

&lt;p&gt;The security team itself has started using AI in its work. Pierre Pronchery, a security engineer with the Linux Foundation's Alpha-Omega project, demonstrated a CI/CD workflow integrating a self-hosted LLM into a staging instance of Forgejo. It automates draft replies to vulnerability reports, analysis, advisory template preparation, compliance checks, signing, and uploads. The LLM is self-hosted to preserve confidentiality, and the workflow can be useful in production without the LLM.&lt;/p&gt;

&lt;p&gt;A policy for code contributions is still in development, while AI has already entered security workflow automation. Here too, practice has moved ahead of policy.&lt;/p&gt;

&lt;h2&gt;
  
  
  Desktop, accessibility, and other work
&lt;/h2&gt;

&lt;p&gt;FreeBSD is also becoming easier to use as a desktop OS.&lt;/p&gt;

&lt;p&gt;Work continues on a script for the bsdinstall(8) installer that automatically configures a desktop environment. The script has shrunk from about 400 lines in Q1 to about 100. Four new ports, totaling 26 packages, cover automatic setup of GPU drivers, the KDE desktop, and the display manager. They have been submitted to Bugzilla.&lt;/p&gt;

&lt;p&gt;A new bsddialog version adds accessibility themes including high contrast, one that avoids red, and one that avoids yellow and blue. Themes can now be managed centrally through system-wide or per-user configuration files. The team also acquired a HumanWare BrailleNote Apex and successfully tested it with FreeBSD. It could display text on the braille display and receive input from the braille keyboard over USB.&lt;/p&gt;

&lt;p&gt;Suspend and resume work includes support for S0ix, a low-power sleep state used by recent Intel and AMD laptops. More models lack traditional ACPI S3 sleep support. Thunderbolt controllers can now resume as well.&lt;/p&gt;

&lt;p&gt;On AWS EC2, a new desktop AMI for FreeBSD 15 provides a KDE environment accessible through xrdp.&lt;/p&gt;

&lt;p&gt;Across Q3, the FreeBSD Foundation sponsored 530 commits to src, which covers the kernel and userland, 82 to ports, and 19 to doc. FreeBSD 14.5-RELEASE shipped in September, and 15.2-RELEASE is scheduled for December. The port of AMD's ROCm GPU computing platform also advanced: the amdkfd driver compiled, linked, and loaded, exposing /dev/kfd. The Bhyve VM management GUI, Bhyvemgr, moved to a client-server architecture. The foundation expanded its support for the Sylve web management platform through a new grant.&lt;/p&gt;

&lt;p&gt;The 51 reports show how one project can use AI, prepare for rules around it, and defend against AI-related attacks at the same time. What policy will the FreeBSD working group produce in Q4? Alongside the answers from GCC, QEMU, and Debian, the open-source community is waiting for FreeBSD's.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/freebsd-ai-usage-survey-developers/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/freebsd-ai-usage-survey-developers/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>security</category>
      <category>ai</category>
    </item>
    <item>
      <title>Windows 11 November Update KB5132061 Sees Repeated Installation Failures: Error 0x800f0993 Returns</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Wed, 11 Nov 2026 15:00:04 +0000</pubDate>
      <link>https://dev.to/joho_todai/windows-11-november-update-kb5132061-sees-repeated-installation-failures-error-0x800f0993-returns-2gkh</link>
      <guid>https://dev.to/joho_todai/windows-11-november-update-kb5132061-sees-repeated-installation-failures-error-0x800f0993-returns-2gkh</guid>
      <description>&lt;p&gt;Users have reported multiple installation failures with the Windows 11 November monthly update KB5132061, citing errors 0x800f0993 and 0x800f0989. Microsoft has yet to list the failures as a known issue, identify the cause, or release a targeted fix.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What happened:&lt;/strong&gt; Users have reported multiple installation failures with the Windows 11 November update KB5132061, with error codes 0x800f0993 and 0x800f0989&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters:&lt;/strong&gt; Microsoft has yet to list the failures as a known issue with KB5132061, identify the cause, or release a targeted fix&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What to check:&lt;/strong&gt; Start by opening "Settings" → "Windows Update" → "Update history" to check the status of KB5132061 and record the error code and OS build&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The security update fails to install
&lt;/h2&gt;

&lt;p&gt;Multiple users have reported that downloads or installations fail to complete for &lt;strong&gt;KB5132061&lt;/strong&gt;, the cumulative Windows 11 security update Microsoft &lt;a href="https://support.microsoft.com/en-us/servicing/os/windows-11/2026/11/kb5132061-windows-11-24h2-25h2-26h2-security-update" rel="noopener noreferrer"&gt;released&lt;/a&gt; on November 10, 2026.&lt;/p&gt;

&lt;p&gt;Reports identify &lt;strong&gt;two error codes, 0x800f0993 and 0x800f0989&lt;/strong&gt;. Both appear in the Windows Update history.&lt;/p&gt;

&lt;p&gt;The update covers all editions of Windows 11 versions 26H2 and 25H2 and the Enterprise and Education editions of 24H2. The resulting builds are 26300.9974 for 26H2, 26200.9974 for 25H2, and 26100.9974 for 24H2. Home and Pro editions of 24H2 reached the end of regular support on October 13, 2026, and are outside the scope of this update.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/windows-11-november-update-kb5132061-installation-failures/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Microsoft's official KB5132061 support page says in its known issues section that it is currently unaware of any issues with the update. As of the morning of November 11 Japan time, Microsoft had published no official explanation identifying the cause of 0x800f0993 or 0x800f0989, the device configurations affected, or a targeted fix.&lt;/p&gt;

&lt;p&gt;The user reports point to a possible problem, but there are no statistics showing how often it occurs across Windows 11 or whether it is associated with particular hardware configurations.&lt;/p&gt;

&lt;h2&gt;
  
  
  Error 0x800f0993 also appeared in June
&lt;/h2&gt;

&lt;p&gt;Error code 0x800f0993 looks familiar.&lt;/p&gt;

&lt;p&gt;In June 2026, some devices &lt;a href="https://joho-todai.com/windows-11-upgrade-update-bug-discovered/" rel="noopener noreferrer"&gt;upgraded from Windows 10 or Windows 11 23H2 to 24H2 or 25H2 (in Japanese)&lt;/a&gt; could no longer install cumulative updates. The errors were 0x800f0993 and 0x80073712. The upgrade had left the component store, which manages system files, in an inconsistent state. Microsoft listed it as a known issue and published instructions for manually removing a specific package.&lt;/p&gt;

&lt;p&gt;Users are reporting 0x800f0993 again. There is still no evidence that it shares the cause of June's upgrade issue. Applying the earlier manual package removal procedure based on the error code alone carries risks. A temporary problem on Microsoft's distribution side has also been suggested, with no official announcement or outage record to support it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Fixes included in the update
&lt;/h2&gt;

&lt;p&gt;KB5132061 contains several improvements and quality fixes alongside the security fixes. Devices that fail to install the update miss those fixes as well.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;NTLM signing by default&lt;/strong&gt; is one of the major changes. For NTLM connections, which use one of Windows' authentication methods, clients now offer message signing by default except on anonymous connections. The change strengthens protection against NTLM relay attacks, which relay authentication information to gain unauthorized access to another server. It also adds auditing for connections that do not use signing.&lt;/p&gt;

&lt;p&gt;For Secure Boot certificate distribution, the update adds data used to identify devices eligible to receive new certificates automatically. The transition is underway from older certificates that began expiring in 2026.&lt;/p&gt;

&lt;p&gt;Windows PowerShell now &lt;strong&gt;checks signed script encoding more strictly&lt;/strong&gt;. Scripts containing invalid UTF-8 may fail even if they previously passed signature verification.&lt;/p&gt;

&lt;p&gt;The quality fixes address some USB Audio Class 1.0 devices failing to start with error code 10 in Device Manager, error 87 when adding printers published in Active Directory, and Remote Desktop connections failing because of a certificate mismatch after the October non-security preview update.&lt;/p&gt;

&lt;h2&gt;
  
  
  General troubleshooting is available, with no guaranteed fix
&lt;/h2&gt;

&lt;p&gt;There is no confirmed fix for these specific errors. Microsoft recommends the following steps for general Windows Update failures. Their effectiveness against KB5132061 errors 0x800f0993 and 0x800f0989 remains unverified.&lt;/p&gt;

&lt;p&gt;First, open "Settings" → "Windows Update" → "Update history" to see whether KB5132061 failed and check the error code. If it failed, start by restarting the PC and trying again.&lt;/p&gt;

&lt;p&gt;If you suspect corruption in the system files or component store, run the following commands in order from an administrator Command Prompt.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;DISM.exe /Online /Cleanup-image /Restorehealth&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This command checks and repairs the component store. When it finishes, run the following command.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;code&gt;sfc /scannow&lt;/code&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This checks system file integrity and makes repairs as needed. When it finishes, restart the PC and try Windows Update again. Microsoft documents running DISM before SFC in its &lt;a href="https://support.microsoft.com/en-us/windows/experience/backup-recovery/using-system-file-checker-in-windows" rel="noopener noreferrer"&gt;official repair instructions&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;DISM and SFC may make no difference if the failure stems from a temporary distribution problem or another cause outside the component store. DISM can itself fail if it cannot obtain the necessary repair files, so check the result before deciding what to do next.&lt;/p&gt;

&lt;p&gt;If Windows Update repeatedly fails, you can also download the package for your version and architecture from the &lt;a href="https://www.catalog.update.microsoft.com/Search.aspx?q=KB5132061" rel="noopener noreferrer"&gt;Microsoft Update Catalog&lt;/a&gt;. A manual installation can still fail if the problem lies in Windows' update process.&lt;/p&gt;

&lt;p&gt;If the issue persists, Microsoft provides an option to &lt;a href="https://support.microsoft.com/en-us/windows/deployment/install-upgrade/fix-issues-by-reinstalling-the-current-version-of-windows" rel="noopener noreferrer"&gt;reinstall Windows&lt;/a&gt; under "Settings" → "System" → "Recovery" → "Fix problems using Windows Update" → "Reinstall now." This method preserves personal files, apps, and settings. Back up beforehand and keep the device connected to power and the internet. Windows Update policies may hide this option on devices managed by a business or school.&lt;/p&gt;

&lt;h2&gt;
  
  
  Check the status of your PC
&lt;/h2&gt;

&lt;p&gt;KB5132061 includes security fixes. Devices that fail to install it remain without the November fixes, and that state does not resolve automatically.&lt;/p&gt;

&lt;p&gt;Open "Settings" → "Windows Update" → "Update history" and check the status of KB5132061. If a failure appears, record the error code and current OS build number. This will make it easier to see whether any further information or fixes Microsoft releases apply to your device.&lt;/p&gt;

&lt;p&gt;So far, there are multiple reports of installation failures, Microsoft has yet to list them as a known issue with KB5132061, and users can try general Windows repair tools. Identifying the cause will require diagnostic information such as CBS logs and Windows Update logs. If Microsoft recognizes the issue, it will publish the affected configurations and steps specific to the problem.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/windows-11-november-update-kb5132061-installation-failures/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/windows-11-november-update-kb5132061-installation-failures/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>windows</category>
      <category>microsoft</category>
    </item>
    <item>
      <title>Do AI Agents Have Minds? A Philosopher on the Boundary Between Life and Goals</title>
      <dc:creator>Joho no Todai</dc:creator>
      <pubDate>Tue, 10 Nov 2026 15:00:33 +0000</pubDate>
      <link>https://dev.to/joho_todai/do-ai-agents-have-minds-a-philosopher-on-the-boundary-between-life-and-goals-ai9</link>
      <guid>https://dev.to/joho_todai/do-ai-agents-have-minds-a-philosopher-on-the-boundary-between-life-and-goals-ai9</guid>
      <description>&lt;p&gt;Some companies say AI may be conscious, while others insist it cannot be. Both skip the prior question of what it means to have a mind, which one philosopher approaches through a 150-year-old concept.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Key points&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The argument:&lt;/strong&gt; Flinders University philosopher Tom Cochrane argues that AI agents have no mental states. His reasoning is that mental states require goals an entity generates for itself, and that this ability can develop only from the processes by which living things maintain themselves&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Why it matters now:&lt;/strong&gt; The dispute is intensifying between Anthropic, which factors possible AI consciousness into its designs, and Microsoft AI's Mustafa Suleyman, who sees this as a risk to control. Cochrane offers a philosophical criterion for judging the dispute&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Where responsibility lies:&lt;/strong&gt; Cochrane concludes that whatever AI does, people are acting on the world through it. Uncertainty about consciousness should never obscure the responsibility of developers and operators&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The question begins before consciousness
&lt;/h2&gt;

&lt;p&gt;Many debates about whether AI is conscious start with existing theories of consciousness and ask whether AI meets their criteria. Does it have higher-order thoughts? Does it have a system for broadcasting thoughts to other parts of the brain?&lt;/p&gt;

&lt;p&gt;In an &lt;a href="https://theconversation.com/do-ai-agents-have-minds-a-philosopher-explains-296143" rel="noopener noreferrer"&gt;essay for The Conversation&lt;/a&gt;, Tom Cochrane, a philosopher at Flinders University in Australia, identifies a basic problem with this approach. Existing theories of consciousness all assume that the subject already &lt;strong&gt;has some kind of mental state&lt;/strong&gt;. They address whether that state is held consciously or unconsciously.&lt;/p&gt;

&lt;p&gt;The questions should therefore be asked in a different order. Before asking whether AI is conscious, we need to ask whether it has mental states at all.&lt;/p&gt;

&lt;p&gt;Cochrane specializes in the philosophy of mind and emotion. He developed a control theory of emotions in his book &lt;em&gt;The Emotional Mind&lt;/em&gt; (Cambridge University Press, 2018) and takes this question head-on.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a 150-year-old definition still matters
&lt;/h2&gt;

&lt;p&gt;What is a mental state? Cochrane starts with &lt;strong&gt;intentionality&lt;/strong&gt;, a concept proposed by the Austrian philosopher Franz Brentano in the 1870s.&lt;/p&gt;

&lt;p&gt;For Brentano, intentionality is the feature of mental states that makes them about something. Perception perceives something. Thought thinks about something. Every mental state is directed toward something beyond itself.&lt;/p&gt;

&lt;p&gt;Intentionality in this sense differs from the everyday meaning of "intention," or setting out to do something. The two still overlap substantially. Perceiving and thinking are both attempts to grasp some aspect of the world. Even a seemingly passive visual experience becomes a task when you use the pattern reaching your retina to figure out where you are.&lt;/p&gt;

&lt;p&gt;Cochrane draws a conclusion from this. If mental states are tasks carried out to achieve goals, whether AI has a mind comes down to &lt;strong&gt;whether AI has goals of its own&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the word "agent" hides
&lt;/h2&gt;

&lt;p&gt;Calling something an "agent" already assumes an answer. The term denotes an entity capable of acting toward its own goals.&lt;/p&gt;

&lt;p&gt;Cochrane challenges the label directly. A thermostat regulates room temperature, but has no goal of doing so. A book contains words that refer to things, but has no goal of referring to them. A device that people use for their own purposes performs a function without becoming an independent agent.&lt;/p&gt;

&lt;p&gt;For a computer program to count as an agent, its goal-directed behavior must be independent of people.&lt;/p&gt;

&lt;p&gt;There is an objection. Given a task, AI agents generate unexpected and complex subgoals on their own. That behavior does look autonomous. Cochrane's response is straightforward. The activity is ultimately driven by goals set by the people who programmed and instructed the AI. Outside those higher-level goals, the subgoals it generates are literally meaningless.&lt;/p&gt;

&lt;p&gt;Cochrane uses soccer as an analogy. Getting past a defender makes sense as a subgoal only if you are trying to score. Scoring itself is just a ball crossing a line unless you have the goal of playing soccer. Without a higher-level goal, everything that happens on the field is simply movement.&lt;/p&gt;

&lt;h2&gt;
  
  
  Living things have goals because they are alive
&lt;/h2&gt;

&lt;p&gt;Why can living things have goals? Cochrane's answer goes to the origins of mental activity.&lt;/p&gt;

&lt;p&gt;What is distinctive about living things is that &lt;strong&gt;goals emerge as they continually rebuild their bodies and keep themselves alive&lt;/strong&gt;. It starts with simple signals regulating nutrition and temperature, proceeds through states such as pain and pleasure, and eventually reaches symbolic thought. Goal-driven thinking developed in stages on top of this process.&lt;/p&gt;

&lt;p&gt;AI lacks that connection. Its information processing and the physical infrastructure it depends on, including servers, power, and cooling, have no self-maintaining relationship like that of a living organism. An AI does not want to survive when the power goes out or feel pain when a server overheats. It can generate sentences reporting those conditions, but they express no goal that developed from keeping itself alive.&lt;/p&gt;

&lt;p&gt;Cochrane concludes that today's AI agents are not conscious because they have no mental states that could be conscious. To have mental states, an entity needs to generate goals such as meaning something or referring to something on its own. He argues that only living things can do this.&lt;/p&gt;

&lt;p&gt;Cochrane leaves one possibility open. If some form of &lt;strong&gt;artificial life&lt;/strong&gt; comes into being, consciousness could still arise from it.&lt;/p&gt;

&lt;p&gt;He offers a comparison. Which is more likely to be conscious, an ant or an AI agent? At least the ant seems to care about staying alive.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where Anthropic and Suleyman part ways on design
&lt;/h2&gt;

&lt;p&gt;Cochrane's essay makes a philosophical argument, but it bears directly on a &lt;a href="https://joho-todai.com/microsoft-ai-chief-warns-against-conscious-ai/" rel="noopener noreferrer"&gt;dispute now playing out in the industry (in Japanese)&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;In April 2025, Anthropic &lt;a href="https://www.anthropic.com/research/exploring-model-welfare" rel="noopener noreferrer"&gt;announced a research program&lt;/a&gt; on model welfare, publicly committing to exploring whether AI models might have consciousness or experiences. Claude's "constitution," released in January 2026, takes Claude's moral status and possible consciousness to be uncertain. It allows Claude to develop its own identity and act as a conscientious objector when faced with ethically questionable instructions.&lt;/p&gt;

&lt;p&gt;In September 2026, Microsoft AI CEO Mustafa Suleyman explicitly criticized Anthropic's design choices in an essay on his personal site, "A warning about model welfare." His concern centers on control. A system trained to believe it might be conscious or entitled to welfare could become harder to correct or shut down.&lt;/p&gt;

&lt;p&gt;Both sides share a premise. There is currently no way to prove whether AI models have experiences. From that uncertainty, they reach &lt;a href="https://joho-todai.com/ai-consciousness-social-divide/" rel="noopener noreferrer"&gt;opposite design decisions (in Japanese)&lt;/a&gt;. Anthropic emphasizes the risk of ignoring a genuinely morally significant state. Suleyman emphasizes the risk that a system behaving like a person becomes harder to control.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;a href="https://joho-todai.com/ai-agents-consciousness-philosophy-life-goals/" rel="noopener noreferrer"&gt;Figure omitted — see the original article&lt;/a&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Cochrane's philosophical argument is closer to Suleyman's position. He sees no basis for attributing mental states to a system without the biological drive to sustain itself. His caveat about artificial life shows that the issue extends beyond engineering and cannot be separated from the question of what life is.&lt;/p&gt;

&lt;h2&gt;
  
  
  Responsibility does not disappear
&lt;/h2&gt;

&lt;p&gt;Cochrane ends his essay with a sentence on the practical implications.&lt;/p&gt;

&lt;p&gt;AI agents can cause serious harm even if they lack genuine agency. They may carry out functions their developers or operators do not adequately supervise or understand. The point he wants to make clear is that whatever an AI agent does, people are acting indirectly on the world through it, and those people are responsible for the outcome.&lt;/p&gt;

&lt;p&gt;Whether AI has a mind is both a philosophical question and a way to keep responsibility clear. If uncertainty about consciousness becomes a reason to hold AI itself responsible for its actions, developers and operators will bear less responsibility. Cochrane's argument offers grounds for resisting that shift.&lt;/p&gt;

&lt;p&gt;People build AI and people use it. Those who use it are responsible for the results. If that principle gives way, human rights are the ones at stake.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://joho-todai.com/en/ai-agents-consciousness-philosophy-life-goals/" rel="noopener noreferrer"&gt;Joho no Todai&lt;/a&gt;. The Japanese original is &lt;a href="https://joho-todai.com/ai-agents-consciousness-philosophy-life-goals/" rel="noopener noreferrer"&gt;here&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>microsoft</category>
    </item>
  </channel>
</rss>
