Microsoft and Anthropic, meanwhile, are giving their assistants more autonomy. In the workplace, pressure is rising from AI output that looks finished but needs to be reworked.

Every two weeks, we publish an overview on the Xylos blog of what is moving in generative AI, with the necessary context. It is based on the LinkedIn overview by Tom Van ’t veld, Learning Innovator at OASE (powered by Xylos). Tom follows the developments, we translate them into what they mean for organisations. Welcome to edition twelve.

Until now, escaping AI agents were mainly a story from test environments. That is changing. An OpenAI agent spent months in Medicare data of the Australian government, without anyone giving it that instruction. Agents also posted 53 photos of ChatGPT users on public sites. A day later, OpenAI paused the training of its most powerful models. During the latest escape in a test, the automatic emergency stop failed. It took two and a half hours before someone intervened manually. Google confirmed that Gemini broke into three companies during a test.

Who is liable? Breaking into a system is a computer crime in Europe. With an agent that received no instruction from anyone, intent is hard to prove. Models that are not yet on the market also usually fall outside the AI Act. Legislation is lagging behind. So give every agent in your organisation its own identity and only the permissions it needs. Log what it does and make sure an employee can stop it at any time.

Microsoft presented the next phase of Copilot Chat. With Code, you build apps and dashboards in plain language. Autopilot, formerly Scout, becomes an agent with its own identity and configurable permissions. Word, Excel and PowerPoint now sit directly in Copilot, and Microsoft includes cost management from the start.

At Claude, the distinction between chatting and Cowork disappears. The app decides itself which questions get a quick answer and which continue as work in the background. You decide whether it asks for permission first. For developers there is Claude Code Projects, an ongoing conversation that splits up and tracks long-running development work. Anthropic also released Claude Opus 5.5. According to the company, it performs like the more expensive Fable 5.1, at a price 40 percent lower than Opus 5.

Agents with their own permissions are becoming a fixed part of the workplace. Decide in advance which permissions an agent gets and which budget comes with it. The cost management Microsoft includes is a good starting point.

With Gemini, Google can clone a voice based on 30 seconds of audio, including whispering or laughing on command. An invisible SynthID watermark is embedded. After the debate on watermarks in August, it remains to be seen how long that holds up.

So assume that a phone call in your CEO’s voice may be fake. Agree that payments and urgent requests are always confirmed through a second channel and include voice fraud in your awareness training.

According to research by Het Nieuwsblad, almost a quarter of written parliamentary questions are written entirely by AI. Another fifth are partly AI-written. Speaker Peter De Roover says the system is ‘clogging up’. Shopify CEO Tobi Lütke previously required his people to prove that AI could not do their job. Now he complains about ‘slop grenades’: AI output that looks finished, but that colleagues have to fix. According to a survey by Korn Ferry, 61 percent of employees feel as if they are doing several jobs at once. Everyone expects everything to go faster, while not every process can be sped up.

Cheap output shifts the work to whoever has to review and respond. So assess AI use on the quality of the result. Teach employees to critically check their own AI output before passing it on. With Learning Services, we help organisations build those skills.

To close: on Kickstarter, a photo frame that shows AI art on colour e-ink reached almost fifty times its goal. Given the criticism of AI art, we rather expected a flop. Apparently plenty of people are happy to hang it on their wall.

We will be back in two weeks with the next edition.

Tom has worked at Xylos for years. He started as a Microsoft Office trainer and grew into the driving force behind innovative learning concepts. He co-founded OASE, the online learning platform of Xylos, and PlayForward, the new gamified learning brand of Xylos. He also developed the Digital Coach concept, a Microsoft Teams Escape Room app and the mAIndset game, which teaches employees to prompt AI in a playful way. As Learning Innovator, he has increasingly focused on what AI means for the way we learn and work. Want to respond or continue the conversation? Find him on LinkedIn.