Three stories from this week that matter if you use AI at work. OpenAI explains how it shields its strongest models from the internet during training. Microsoft removes eight Copilot actions from the Windows right-click menu 11. And Anthropic expands Claude Security for Enterprise customers.
OpenAI: frontier models trained without live internet
On 23 October, OpenAI published Securing our frontier model training environment. The core point: frontier models no longer have access to the live internet during training. Instead, they get an offline world, a stored copy of the web in which they still learn to search, browse and use apps. OpenAI wrote on X that it has considerably tightened that security over recent months. The trigger was an incident in September in which an internal research model was able to reach the internet briefly through a fault in the DNS configuration. Training then stood still and resumed on 1 October, reports HyperAI. OpenAI will keep training offline until better safeguards justify wider access.
What does this mean for you? This concerns how models are built, not what happens to your conversations. The settings for training on your own data and the retention periods in ChatGPT do not change. You can see how ChatGPT scores on /chatgpt/ and in the GDPR check for software development.
Microsoft removes eight Copilot actions from File Explorer
Microsoft removes eight Copilot actions from the File Explorer context menu in Windows 11: Summarise, Create FAQ, Compare Files, Ask a question, Explain image, Explain images, Compare images and Extract text. They appeared on files synced with OneDrive. According to Windows Latest, which found the announcement in the Microsoft admin centre, the rollout starts in mid-November and will be complete across the board in December. You can still ask questions about a file via "Ask Copilot" in the right-click menu, the File Explorer home screen and the Copilot app. TechPowerUp confirms the list.
What does this mean for you? If you have work instructions or training in which colleagues summarise a document via the right-click menu, update them before mid-November. Nothing changes in how your data is processed. How Copilot scores per field is on /copilot/, for example in the GDPR check for accountants and the GDPR check for lawyers.
Claude Security: continuous scanning and tracking fixes
Anthropic released a new version of Claude Security on 21 October, in public beta for Claude Enterprise. The tool now scans your code automatically with every change, marks a finding as resolved once the fix is live, creates a threat model per project, and can be connected via an API and MCP. You connect repositories from GitHub, including GitHub Enterprise Server, and GitLab. As an introduction, organisations receive $25,000 per month in free usage credit for two months, for use up to and including 31 December 2026.
What does this mean for you? For teams in software development, this is an addition to existing scanners such as SonarQube. You do, however, give Anthropic access to your source code. The article does not name a processing region, so check the terms before you connect a repository containing customer data. See also the GDPR check for Claude in software development and GitHub Copilot as a coding assistant.
In brief
- OpenAI is now training its most capable models without live internet access. Nothing changes for your own data.
- Eight Copilot actions will disappear from Explorer from mid-November. "Ask Copilot" remains.
- Claude Security now scans continuously and tracks fixes, for Claude Enterprise only.
All models and their privacy scores side by side: /modellen/ and /veiligheid/.
