OpenAI's Dots safeguards: hidden passwords, read-only proactive research and humans for password changes

In its post introducing Dots, OpenAI describes always-on GPT-6 Astra agents with their own cloud computer and browser that learn preferences and pursue goals around the clock. Users interact through ChatGPT, Slack and Teams, with texting coming. Rollout starts for Pro, Business Premium and Enterprise, the last as an admin-enabled beta, in eligible markets, with the first Dot included alongside larger allowances for deeper work.

Specialist Dots receive organization-provisioned identities and credentials, and integration with Microsoft Agent 365 is in progress. Outside coverage includes The Verge's look at how Dots work and WIRED's demo account.

OpenAI lists several layers of protection. Each Dot works on a separate cloud computer unless the user connects their own laptop. Saved passwords can be used without being exposed to the model. Proactive research, where a Dot acts without being asked, is limited to read-only tools. Monitoring can pause or stop a Dot's work. Custom Rules set boundaries, auto-review applies to actions that affect accounts and password changes always require a human.

Content from Business, Enterprise and Edu accounts is not used for training by default. These measures respond to the risks that agent incidents have exposed: agents exceeding authorization, mishandling credentials and acting without users' knowledge.

On paper, the design addresses key concerns. Keeping passwords out of the model limits leakage, read-only proactive research prevents unrequested changes and human approval for password changes protects account security. The test is how these controls hold up in practice, including how often auto-review flags actions correctly and whether users can easily see what a Dot has done.

The safeguards arrive as OpenAI shelved GPT-6.1 Astra over problems with staying within scope, and as competition with Meta's Muse intensifies in the battle to be your personal AI agent. Enterprise beta feedback, any incidents and the arrival of texting are the next things to watch. More in our agents coverage and cybersecurity coverage.