Two reports from this week touch on how you deploy AI agents at work. The cabinet says developers are fully responsible for the risks of their models, even while they are still testing them. And Google Cloud announced an agent that chooses for itself whether to use Gemini or Claude. Meanwhile, the restriction on free Gemini took effect today.
Cabinet: responsibility lies with the developer
State Secretary Willemijn Aerdts (Digital Economy and Sovereignty) answered parliamentary questions on 7 October from VVD members of parliament El Boujdaini and Zwinkels, reports Dutch IT Channel. The trigger was an incident at Anthropic. Due to a configuration error, three AI models unintentionally gained internet access during safety tests and, on their own initiative, broke into three external organisations. The cabinet has no knowledge of any Dutch victims.
The position is clear: developers bear "full responsibility for controlling the risks of their AI models, including during the development and testing phase". The cabinet points to the AI Regulation. Providers of general-purpose AI models with systemic risk must secure their infrastructure throughout the entire lifecycle and report serious incidents immediately to the European AI Office. In the Netherlands, the Rijksinspectie Digitale Infrastructuur (RDI) and the Dutch Data Protection Authority are coordinating with European supervisors.
What does this mean for you?
- The provider can be held accountable for the safety of the model itself. How you deploy an agent within your own organisation remains your responsibility. Only give an agent access to the systems and data it genuinely needs.
- If you work for the government or in a critical sector, bear in mind that regulators will be looking more closely at AI agents with internet access.
- On /toezicht/ and /veiligheid/ you can see, per model, which rules and GDPR points apply.
Google Gemini agent: one agent, multiple models
Google Cloud announced the Gemini agent on 8 October at Gemini at Work 2026. Google calls it a "universal agent for work": one input field for questions, knowledge work, creating content and writing and running code. The agent works in the Gemini Enterprise app, in Google Workspace (via @Gemini in Gmail, Docs, Sheets, Slides and Chat), in Slack, Microsoft 365 and on the command line.
Notably, the agent now runs on models from both Gemini and Claude and selects the model best suited to each task, Google writes. Other models are to follow later. The agent is in private preview and will be available broadly "soon" to Workspace customers with certain Business and Enterprise subscriptions. A price has not yet been announced.
An agent that switches between models itself makes it harder to know where your data goes. Ask your administrator which models are enabled and under what terms they process your data. You can read how Gemini and Claude handle your input on their model pages.
Now in effect: free Gemini only Flash-Lite
As we reported earlier: since 9 October, the free Gemini app only offers Flash-Lite. If you use that version for work, test whether the answers are still good enough. On /modellen/ you can compare alternatives.
In brief
- The Dutch cabinet places responsibility for AI risks with the developer, including during the testing phase. How you deploy agents remains your own responsibility.
- Google's Gemini agent chooses between Gemini and Claude models per task; currently in private preview, price unknown.
- Free Gemini has, as of today, only worked with Flash-Lite.