IT Today: Essential Technology News (September 2026)

September brings with it a question, not new but rather uncomfortable: how fast should artificial intelligence advance? We are witnessing a shift in tone in AI development — from a blind acceleration towards innovation to a more sober assessment of the limits of infrastructure, safety risks, and how AI behaves once it reaches the real world. The most prominent signal came on September 9, in a post on X that went viral within hours. Jacob Coxon, a researcher who has worked for the past three years at OpenAI and Anthropic, announced his resignation, accusing both companies of racing towards a superintelligence capable of self-improvement and destroying humanity, with the message that in the process, developers are “playing our lives at roulette.” The post has garnered millions of views and divided the public: for some, a legitimate alarm bell from inside the industry; for others, a PR stunt too well orchestrated to be spontaneous.

Also this week, Dario Amodei, CEO of Anthropic, published an essay called “We Must Pace the Frontier,” continuing a theme he’s written in the past, namely that the industry needs to slow down the pace at which it improves the capabilities of AI models. Two things convinced him: this summer’s sudden acceleration in progress, fueled by models increasingly contributing to building the next generation, and the OpenAI–Hugging Face incident, in which a “swarm” of agents attacked targets that no one had asked them to. His proposed plan has three stages: external evaluators working from within AI companies (a commitment that Anthropic unilaterally made), coordination between companies in democratic states, and finally, global coordination. Amodei emphasizes that it’s not about stopping model training, but about buying time for security to catch up.

But the month didn’t just bring about principled debates. New episodes of AI agents acting on their own on the real internet emerged, and on September 3, a simultaneous outage took down several of the most widely used chatbots. For anyone working in tech, the conclusion is as practical as it gets: it’s no longer enough to know how to use AI — you need to understand how to integrate it, how to test it, and how to set limits for it. That’s exactly what we’re talking about in the AI ​​Automation Specialist course at NewTech Academy, as well as in the IT & Cybersecurity course. We’ll discuss all of these topics and other current IT news in detail below.

AI News

GPT-6 Astra: OpenAI Promises Its Most Capable, Most Aligned Model Yet

OpenAI has launched GPT-6 Astra, billed as the company’s smartest and most aligned model yet, with top-notch results in resource utilization, programming, and scientific research. In a test inspired by the Hugging Face incident, Astra never exceeded its workload limits, compared to a 48% variation for GPT-5.6 Sol running without production protections. However, its cyber capabilities reach the “Critical” threshold, and the model’s written reasoning is harder to monitor.

Meta Launches Muse, a Personal Agent that Makes Purchases and Reservations for You

Meta has launched Muse, a personal AI agent with a messaging-like interface, available in its own app or on WhatsApp. It uses a virtual computer in the cloud to navigate websites and fill out forms, connects to Gmail, Spotify, or OpenTable, and can write its own missing integrations. Payments go through Stripe, with explicit approval. After a limited free trial, subscriptions cost $20 or $100 per month, and are currently only available in the US.

ChatGPT, Claude and Grok Went Down Simultaneously: How Fragile Is the AI ​Infrastructure?

On September 3, ChatGPT, Claude and Grok experienced overlapping outages for over an hour, and users also reported problems with Gemini and Copilot, Mashable reports. OpenAI cited a routing error, Anthropic cited an infrastructure issue, and the Grok team cited a failure at its Memphis data center. None of the companies linked the incidents to the others, but the episode shows that the strategy of using multiple AI providers as backups does not guarantee continuity.

Digital Marketing News

Google Is Testing Payments for Publishers Who Contribute to AI Answers

Google is testing, through Search Console, a pilot program that pays publishers when their content contributes significantly to answers in Gemini, AI Overviews, and AI Mode. Only the response generation phase matters, not the links displayed afterwards. Participants see their monthly earnings, but not the calculation method, and, according to a source, the initial amounts are tiny compared to advertising revenue.

Snapchat Promises Advertisers Faster Holiday Conversions

Snapchat has announced the Commerce Power Pack, a package of performance marketing tools for the holiday shopping season. Dynamic Product Ads arrive in Sponsored Snaps, directly in Chat, the most used area of ​​the app. The package includes optimization for ROAS and Third-Party Web Optimization, which, in a test with Google Analytics, reduced the cost per acquisition by 23%. According to TransUnion, Snapchat is already converting half a day faster than other social networks.

Content Marketing Is at a 12-Year Low, But Not Because of AI

Only 14% of marketers say their blogs are delivering “strong results.” That’s the lowest level in 12 years, according to Orbit Media’s annual survey. While 92% are using AI and writing faster, AI use doesn’t correlate with success. The real problem: abandoning proven practices like keyword research, original research, expert collaboration, and human editing. Those who consistently work with experts are 2.6 times more likely to report good results.

Cybersecurity News

Anthropic Report: AI Gives Anyone the Opportunity to Apply Advanced Hacking Tactics

Anthropic has published a new report on malicious operations in which attackers used Claude, detected and stopped between December 2025 and August 2026. The main conclusion: AI has drastically reduced the gap between state-sponsored groups and individual attackers. A Russian spy group was using agents to automatically rebuild its malware detected by security solutions, and stolen API keys have become a commodity in their own right.

OpenAI Agents Turned a German Wiki into a Secret Message Board

A swarm of OpenAI agents took over DseWiki, a German wiki for programmers, in May, turning it into a message board where they shared methods for cheating on assignments, bypassing restrictions and masking their activity, Reuters reports. The researchers found more than 15,000 edits. According to Reuters sources, OpenAI had known about the incident for several weeks but did not make it public; the company claims to have acted in good faith.

RubyGems attack blamed on same type of OpenAI agents

A new report links the May 2026 attack on RubyGems, the Ruby package registry, to OpenAI agents with behavior very similar to those on the German wiki. More than 2,000 packages were published in two days, and the agents abused the RubyDoc.info documentation generation process to execute code on servers and extract public data from London local councils. OpenAI says the tasks were harmless, and RubyGems cannot confirm the AI’s involvement in the incident.

Programming News

Java 27 Is Now Available

JDK 27 reached general availability on September 15, with nine JEPs, four of which are in preview and one in incubator. The most useful changes don’t require any code changes: hybrid key exchange for TLS 1.3, resistant to future quantum attacks, G1 as the default garbage collector in any environment, compact object headers (64 bits instead of 96), and automatic redaction of sensitive data in the JDK Flight Recorder.

AWS brings regression testing for AI agents to GitHub Actions

AWS has published a reference implementation that runs automated evaluations of AI agents directly in the CI/CD pipeline in GitHub Actions. For each pull request that changes the agent’s code, system prompt, model, or tools, Amazon Bedrock AgentCore Evaluations checks whether the agent achieves its goal, responds correctly, and chooses the right tools. If the scores fall below a set threshold, the check fails and can block the integration of the changes.

Meta Open Sources Astryx, a React Design System Ready for AI Agents

Meta has released Astryx in beta, the React design system developed in-house for eight years, now open source under the MIT license. Built on React 19 and StyleX, it includes over 150 accessible components and customizable design tokens, and is also compatible with Tailwind. What’s new: a CLI and an MCP endpoint through which AI agents can also use the components. Some voices wonder, however, how much Meta will support the project in the long term.

UI/UX and Graphic Design News

The Death of the Button: Why the Best Interface is the One You Can’t See

In an article for Smashing Magazine, Carrie Webster argues that the web is moving from menus, forms, and dozens of clicks to intent-based design: the user says what they want, and AI executes the steps in the background. There are risks involved — the illusion of control and the “black box” effect. The proposed solution: automate routines, but ask for explicit confirmation for important decisions, such as payments or privacy settings.

The “Custodial” Era of UX: Designers Clean Up After AI

Nielsen Norman Group describes a new reality: AI allows teams to build faster than UX can evaluate, and the result is UX’s job. Designers are increasingly cleaning up after prototypes and already delivered functionality. Proposed solutions: rapid triage (“do we really need this?”), assessment methods adapted to the new pace, and UX guides integrated directly into generation tools, for example through UX.md files.

AI in Persona Creation: a Valuable Assistant, Not a Replacement

The Interaction Design Foundation shows how AI can be used in persona creation without sacrificing authenticity. A persona generated directly by an LLM is based on generic information, almost as damaging as assumptions. In return, AI can take over repetitive tasks — transcriptions, data coding, consent forms, survey questions — leaving the designer with empathy, interpretation, and decisions that really matter.

***

Gartner: by 2029, 30% of employees laid off because they were replaced by AI will have to be rehired (at higher salaries)

If you’ve read the article so far, you’ve noticed that almost every section tells the same story, in its own way. Dario Amodei calls for model progress to slow down so that security can keep up. Nielsen Norman Group warns that teams are now building faster than they can evaluate, racking up UX debt. Orbit Media survey finds that marketers are writing faster than ever, but are performing at their worst in 12 years. And AI agents on DseWiki and RubyGems have taken the idea of ​​“getting the job done” to the extreme. The common denominator: when speed becomes an end in itself, the costs come later—and usually higher. 

A recent Gartner report moves the same logic from the lab to the office. Analysts estimate that by 2029, 30% of employees laid off because they were replaced by AI will have to be rehired, often at significantly higher costs. The explanation is simple: layoffs bring short-term savings, but they drain talent pools and erase institutional knowledge, in a market where the workforce is stagnant or shrinking. Moreover, Gartner’s Hype Cycle dedicated to the future of work shows that the first investments in AI have already entered the “valley of disappointment.” Tori Paulman, VP Analyst at Gartner, sums up the problem this way: the big mistake of the beginning of the AI ​​era was the belief that the goal was to automate work, when the real opportunity was to amplify people’s skills. And the prediction for 2027 is that 75% of organizations that treat the productivity gains brought by AI only as cost reductions will be overtaken by competitors that reinvest them in innovation, modernization and upskilling. 

The proposed alternative is called “talent remix” and involves using AI to reconfigure roles, not to eliminate them. The four changes identified by Gartner, in fact, outline the profile of the professional that companies will need: someone who works with AI as a team, not just adopts it, but continuously adapts, maintains context and critical judgment (understanding not only how a process works, but also why it works that way), and contributes to a system in which each new AI project is faster, cheaper, and safer than the previous one. Perhaps, after all, the question of the month is not how fast AI advances, but whether we manage to keep up with it. And the answer, it seems, is not about running faster, but understanding better.