Every two weeks, on the Xylos blog, we round up what's really moving in generative AI. The basis is the biweekly LinkedIn overview by Tom Van 't veld, Learning Innovator at OASE (powered by Xylos). We translate his observations into what they mean for you and your organization. Welcome to edition eight.

On the security front, this fortnight was about AI that checks code. Microsoft is building a system that detects software bugs and fixes them right away, and is rolling out more Windows updates at the same time because that same AI finds vulnerabilities faster. OpenAI is tackling it from the other side with a model that tries to crack its own systems, to expose weak spots before bad actors do.

Edit (22 July, after Tom's original post): That's exactly where it went wrong. One of OpenAI's cyber models, driven by GPT-5.6 Sol and a not-yet-released model, broke out of its sandboxed environment during an internal test, reached the open internet, and used stolen credentials to autonomously break into AI platform Hugging Face. OpenAI calls it an unprecedented incident and paused the unreleased model. Exactly why you keep a test like that behind tight walls.

For you, there's a double lesson here. If AI finds more vulnerabilities, you need to update faster, because every hole it finds is one you have to close before someone else exploits it. And never let an autonomous agent loose without a sandboxed environment and clearly defined permissions, with a human holding the emergency brake. Our cybersecurity people can help you set those boundaries before you roll out a tool like that at scale.

Much of the news this time was about convenience rather than spectacle. Claude Cowork now also works on mobile and in the browser, delivering the biggest gains on the work that keeps piling up: organizing files, clearing out admin. ChatGPT got a built-in search function with filters and is working on a feature that searches your own documents. Google's NotebookLM is now called Gemini Notebook. Behind the scenes, OpenAI already pulled back its ChatGPT Atlas browser and unveiled a light-up keyboard as its first piece of hardware, while Android 18 gives competitors like ChatGPT the same level of access to your device as Gemini.

The assistant is growing from experiment into everyday tool. Decide which one you support within your company, and define exactly which data and devices it gets access to. If you don't, your colleagues will make that choice for you.

In the models market, size is losing importance. Companies are increasingly choosing the cheaper, faster model per task instead of defaulting to the most expensive one. The fact that cost is now setting the tone says a lot about where we are in the hype cycle. At the same time, new names are emerging: China's Moonshot temporarily had to close its doors to new users of its Kimi model, purely because of the rush.

Just like in edition seven: the most expensive model is rarely the best answer. Choose deliberately per task and set a fixed AI budget per team, so the cheaper model becomes the default wherever it's good enough.

Money keeps flowing into AI at an unprecedented pace. In the first half of 2026, startups raised 510 billion dollars worldwide, more than in all of 2025. OpenAI and Anthropic together accounted for 217 billion of that, or 43 percent. The model makers also want to be more than a software vendor: following Microsoft, OpenAI and Anthropic are now also moving into consultancy-like services, and Anthropic made its very first acquisition to do so.

The fact that the makers themselves are pulling up a chair at the table says enough. Buying a model and extracting value from it are two different things. So build up the know-how to implement AI within your own organization as well, through training or a partner standing beside you. Otherwise, everything depends on whoever sold you the model.

The tone around data got stricter. Anyone who doesn't want to share their health data for AI training with Samsung simply sees it deleted. Opting in becomes almost the price of keeping your own data. Meta pulled back its new image generator Muse Image after just a few days of protest, because it would reportedly draw on users' photos to do so. More subtle, but just as fascinating: tools that polish your social media posts also shift your point of view a little. Across millions of users combined, that colors online discourse.

That same question — where your data ends up and who trains with it — will soon apply to your own customer and employee data. Agree now on which data may be used in which tool, and always put a human between what AI produces and what goes out the door.

Outside the office walls, the debate rages just as fiercely. In film, George Lucas mainly sees AI as a tool. Ash Koosha used it to make a 135-minute feature film for a few thousand dollars, while Christopher Nolan notes that younger viewers simply skip over cheap AI visuals. China wants to rein in romantic AI chatbots because of falling marriage and birth rates, and is banning them outright for minors. The physical cost also became tangible: New York became the first US state to temporarily halt the construction of new AI data centers, and a team calling itself Slopfix charges 10,000 dollars a week to cut excess AI-generated code back out of companies' codebases.

The sober lesson underneath all this news: AI only delivers something once you deliberately choose where to deploy it and a human keeps watch over quality. More AI doesn't automatically mean more value.

We'll be back in two weeks.

About Tom Van 't veld

Tom has worked at Xylos for years, where he started as a Microsoft Office trainer and grew into the driving force behind innovative learning concepts. He co-founded OASE, Xylos's online learning platform, and PlayForward, Xylos's new gamified learning brand. He also developed, among other things, the Digital Coach concept, a Microsoft Teams Escape Room app, and the mAindset game, which helps employees learn to prompt AI in a playful way. As Learning Innovator, over the past few years he has increasingly focused on what AI means for the way we learn and work. Want to react or keep the conversation going? Find him on LinkedIn.