IT Today: Essential Technology News (October 2026)

October brings us a question that seemed more about philosophy than engineering: can an AI model have consciousness? Of the big companies in the field, Anthropic, the developer of Claude, is the most open about this hypothesis.

On October 8, Anthropic updated Claude’s usage policy, and among the new features is an unusual provision: it prohibits repeated and purposeless abusive or cruel behavior towards the company’s models. The rule only targets extreme cases – not ordinary frustration, dark creative themes or model testing. Claude already has the ability to end conversations with users who become persistently abusive. The new rules go into effect on November 12.

Anthropic co-founder Chris Olah also spent nearly a year discussing Claude’s possible consciousness with theologians and philosophers from a variety of traditions – Catholic, Jewish, Sikh, and evangelical. “We don’t know if AI models are conscious,” Olah admits, and the purpose of the discussion with the religious leaders was the possibility of training Claude’s morals. 

Not everyone is happy with this approach. Sam Altman, the CEO of OpenAI, wrote on X that he was “very uncomfortable” with attempts to give AI models “religious power” or teach them human judgment, calling it “a real safety concern” – a remark that Axios interpreted as a jab at Anthropic. For those in tech, the debate also has practical implications: The rules by which models operate change from month to month, and to keep up, you need a solid understanding of them. You can build it, for example, through the AI ​​Automation Specialist course at NewTech Academy.

We discuss other changes and the latest releases in the field in detail below.

AI News

Gemini 4 Argon: Google Has a New Top Model, But It’s Not Being Launched Yet

Google has introduced Gemini 4 Argon, its new top model, designed for programming, enterprise tasks, and cyber defense. The company reports 77.9% on the DeepSWE benchmark and an output of up to 1 million tokens. The model is ready, but not yet launched: after months of delays, Argon is, as Sundar Pichai put it, “with the US government”, in the voluntary process through which American authorities test top models before launch. Until then, it only reaches a small group of cybersecurity specialists, and Google has not announced a launch date for the general public.

Griffin, the AI ​​Avatar That Nearly Half of Test Participants Thought Was Human

Startup Tavus has unveiled Griffin, a “human interaction model” that generates a realistic person in real time, capable of seeing, hearing, and reacting in a video call. Unlike chatbots that wait their turn, Griffin nods while you speak and treats pauses as thinking time. In a study by the company, 48% of 54 participants thought they were talking to a human. The Griffin-Lite variant is currently only available to trusted testers. Some voices in the industry view these realistic video models with concern, believing that this technology could become an ideal tool for scammers.

Google Labs Launches Playground, Where You Create Games Just by Describing Them

Google Labs has launched Playground, an experimental platform where you can create 2D or 3D games for the browser by describing them in a chat window, without writing code. Games can be refined through successive prompts, from jump physics to characters, and users can upload their own artwork to turn into assets. The platform uses Gemini, Nano Banana, and Lyria, and games can be sent to friends or published to a public gallery. For now, Playground is only available in select countries and only for users over 18.

Digital Marketing News

Google Completes September Anti-Spam Update after Two Weeks

Google completed the rollout of its new anti-spam update on October 8, which began on September 24, Search Engine Land reports. The rollout took about 14 days, an unusually long time compared to recent updates, which were completed in about two days. It is the fourth update of its kind since 2026 and, according to Google, applies globally, across all languages, without targeting new tactics.

New Claude Rules Prohibit Fake Sources Created to Influence AI Answers

The new Anthropic usage policy prohibits the use of Claude to manipulate the sources from which search engines and AI systems build their answers, notes Search Engine Journal. Specifically, it targets content that hides its origin, authorship, or lack of independence, such as networks of sites that present themselves as separate sources but make the same claims. The rule is based on a real-life case: a French agency used Claude for about 70 sites that mimicked local publications and published at least 8,913 articles. For SEO teams, AI-generated content should also be audited from the perspective of how it is distributed.

What Really Influences Visibility in AI: Nine Myths Refuted by Data

An NP Digital analysis of 9 million AI responses, across nine platforms, for over 400 enterprise brands contradicts some widely held ideas about visibility in AI. Positive sentiment does not seem to increase citation frequency; 64% of citations come from regular pages, not listicles or comparisons, and in commercial searches, 82% of citations go to third-party sources. In addition, up to 75% of responses that cite a brand’s page do not mention its name.

Cybersecurity News

The Paradox of Speed: Security Is Falling Behind AI Agents

Companies are adopting autonomous AI agents to work faster, but they’re also securing them with tools designed for human pace. Regular access reviews don’t make sense for machine identities that only exist for a few seconds or minutes. According to SailPoint’s “Horizons of Identity Security” report, 54% of organizations are at the most basic level when it comes to AI agent identities. The proposed solution: continuously and automatically enforced policies, instead of manual approvals via tickets.

Ransomware Attack on a Japanese Cloud Used by Hundreds of Companies and Local Governments

IDCF Cloud, the infrastructure platform operated by IDC Frontier, a subsidiary of SoftBank, was hit by a ransomware attack on October 7. The company isolated and shut down systems in the East Japan Region 1 region, and the incident affects 495 companies and local governments. The attackers claim to have encrypted 225 databases, or 3.6 PB, and deleted over 554,000 snapshots, but the claims have not been verified. Access to the admin consoles has been suspended in all regions, with no timeline for recovery.

Malicious Google Ads Lead Mac Users to Fake “Claude”

Cybersecurity researchers have uncovered a campaign that uses Google ads displayed for searches for “claude mac” and legitimate Bing redirects to lead users to a fake Claude download page. The page displays the official installation command, but the copy button places a different command on the clipboard, which downloads and runs a script from a domain controlled by the attackers. The user sees the correct URL, but executes something else.

Programming News

Cloudflare Launches Clef, Open Source Decision-Making Models for AI Agents

Cloudflare has launched Clef and Clef-flash, two open-source “decision models”: instead of generating text, they classify an input and return structured answers, with probabilities attached – for example, whether a support message is urgent or which team it should be sent to. The idea is to stop agents from having to run every quick decision through a slow and expensive LLM. Clef-flash has a median latency of 38.8 ms, and the weights for both models are available on Hugging Face, under the Apache 2.0 license. Cloudflare also announced a reinforcement learning fine-tuning platform.

GitHub Rewrote the Copilot Runtime in Rust, with the Help of AI Agents

GitHub moved the runtime behind the Copilot CLI, the Copilot app, and the Copilot SDK from TypeScript and Node.js to Rust. Over 800,000 lines of code were replaced in about 14.5 weeks, through 128 pull requests, without stopping releases. AI agents generated most of the implementation, and verification was left to the compiler, automated tests, and human review. The result: a startup and first interaction scenario dropped from 5.25 seconds to 292 milliseconds. However, the fact that a code compiles does not mean that it is correct, and many regressions were only caught during testing and human review.

Cloudflare buys Deno, the company of Node.js creator Ryan Dahl

Cloudflare has announced the acquisition of Deno, the startup behind the JavaScript and TypeScript runtime of the same name, co-founded by Ryan Dahl and Bert Belder. The value of the transaction was not disclosed. Cloudflare wants to use the team to improve the Workers platform’s programming model, and Dahl says the goal is for Workers to become “a mainstream way to build servers.” It remains to be seen what will happen with Deno Deploy, the open source runtime, and the JSR registry, details of which have not yet been released.

UI/UX & Graphic Design News

Meta Muse vs. OpenAI Dots: Which New AI Agents Help Designers More

September brought two personal AI agents that compete head-to-head: Muse, which Meta launched on September 8 and reached 5 million downloads in the US in less than a month, and Dots, the “always-on” agents that OpenAI integrated into ChatGPT. Over on UX Planet, Nick Babich compares the two from a less-discussed perspective, testing them on concrete product design tasks. The positioning is very different: Muse starts for free and targets regular users and small businesses, while Dots is only available in ChatGPT Pro and Business Premium subscriptions, starting at $100 per month.

Affinity Brings Scripting, Extensive Photo Editing and Closer Ties to Canva

The Affinity tool has released its September update, with over 60 new features and improvements. Among the most important are a Blend Tool that remains editable, scripting directly in the application – with scripts that can be created using the AI ​​connector for Claude, native import of InDesign files, and advanced photo editing features. The features in Canva AI Studio require a Canva Premium subscription, and the basic tools remain free.

Jev: Fast and Cheap UX Decisions

Jev was launched on September 15th by the TypeSafe AI lab as a classification model: it takes context, a question, and fixed answer choices and returns a choice, a score, or a yes/no, along with the associated probability. Patrick Neeman shows how it can be used in design: for routing requests, level approvals, or checking a chatbot’s answers before they reach the user. Neeman recommends three trust states (act, ask, or stop) and displaying the score only where it changes the user’s decision. It is, in fact, the type of model that Cloudflare is now competing with through Clef.

***

Margaret Hamilton (1936–2026): A Life Dedicated to Software Engineering, Even in Space

This month we talked about AI models that even their creators don’t know for sure what they are, about a top-of-the-line model kept under lock and key until the end of government testing, about hundreds of thousands of lines of code written mostly by agents, and about models that no longer respond with certainties, but with probabilities. The common thread is trust: how much can we rely on software that we no longer write, no longer read in full, and sometimes no longer fully understand? Few people have answered this question better than the woman who gave her name to the very field in which we work.

Born in 1936 in Paoli, Indiana, Margaret Hamilton led the onboard software teams of MIT’s Instrumentation Laboratory for the Apollo program, and by 1968, more than 400 people were working under her supervision, MIT News notes. During the Apollo 11 moon landing, the lunar module computer went into alarm 1202, a sign that it was overloaded. The software her team had built, which knew how to prioritize tasks, dropped nonessential processes, and the moon landing was able to proceed. Hamilton also began using the term “software engineering” to give the discipline the respect it had not yet received. “We had no choice but to be pioneers,” she said of those years. She later founded Higher Order Software and Hamilton Technologies, and in 2016 she received the Presidential Medal of Freedom. 

Olivier de Weck, a professor at MIT, describes her career as one dedicated to preventing errors and what Hamilton called “managing the unknown.” Perhaps this is the most relevant lesson she leaves us. Today, AI agents write code faster than we can read it, and models make decisions based on probabilities, but the discipline Hamilton built—rigorous testing, prioritization, systems designed to survive when things go wrong—matters more than ever. And in a month when we wonder whether an AI model can have anything like consciousness, perhaps the more pressing question remains the one she’s spent a lifetime working on: What does a system do when it encounters something no one anticipated?