Three items from the past few days. Early in 2027, the cabinet presents goals for AI in central government; with Gemini you can now edit generated files yourself; and a report about a Claude agent on a US defence network shows what goes wrong when you give an AI agent your own login details.
Cabinet: AI goals for the civil service in the first quarter of 2027
On 26 October, Eric van der Burg, State Secretary for Kingdom Relations and Efficient Government, sent a progress letter to the House of Representatives. According to Rijksoverheid.nl, 2027 is to see €900 million saved on central government, rising to €3.5 billion from 2035. That amounts to an estimated 25,000 fewer FTEs, spread over several years. In the first quarter of 2027, the cabinet will present goals and savings plans for using AI to work more productively.
The letter itself is cautious about AI. iBestuur quotes it as saying that the effects of AI and automation on employment should not be overestimated, as they also led to new tasks and more outsourcing. The 25,000 FTEs therefore do not come from AI alone, but also from fewer regulations, standard ICT and less office space.
What does this mean for you? If you work for the government, the question of which AI you may use in 2027 becomes much more concrete. In the meantime, compare the GDPR check for Copilot in government, the GDPR check for ChatGPT in government and the professional tools at AI for government.
Gemini: edit created files directly
On the 29 of October, Google announced that you can now edit files that Gemini creates for you within the app itself. This covers PDFs, Google Docs, Google Sheets, presentations, Markdown, text files and code. You open the file full screen, adjust it or select a passage of text, and ask Gemini to rewrite only that part. PDFs work on the web and mobile, the other formats on the web only. According to Google, it is available to all users, in all languages in which the Gemini app is available.
What does this mean for you? You will need to copy drafts into Word or Docs less often. If you upload documents containing client or patient data, first check your account and settings on the Gemini page, for example via the GDPR check for Gemini in education. If you are torn between two assistants, see Claude vs Gemini.
Claude agent ran for two months on a Pentagon network, according to WIRED
On 29 October, WIRED reported, citing anonymous sources, that an employee working for Defence through Accenture connected a Claude Code agent on a personal account to a development environment on a US Department of Defense network. According to The Next Web, the agent thereby obtained that employee's login credentials and administrator rights, and ran for two months before anyone noticed. Accenture says the allegations were independently investigated and found to be unfounded. The Pentagon did not respond, and TNW has not been able to verify the story itself.
What does this mean for you? What exactly happened is therefore not yet established. The lesson is clear, however: never give an AI agent your own passwords or administrator rights, and do not use a personal account for work outside the view of your IT department. This also played a part in an unapproved AI tool at a US hospital. If you programme with AI, look at the GDPR check for Claude in software development, GitHub Copilot and AI for software development.
In brief
- The cabinet wants 25,000 fewer FTEs in central government and will present AI goals in the first quarter of 2027.
- In Gemini, you can now directly edit PDFs, documents and code that Gemini creates.
- A Claude agent reportedly ran for two months on a Pentagon network, Accenture denies it. Never give agents your own login credentials.
- All models and their privacy scores are at models and safety.
