Article

How we build safety, security, and privacy into dots

OpenAI · openai.com

How an always-on agent is fenced in: its own sandboxed computer, read-only background research, auto-review of actions and approval for purchases.

Open on openai.com ↗

It answers 4 questions

  1. What should I hand to an AI agent, and what should I keep doing myself?
  2. How do I check what an AI agent actually did before I trust its work?
  3. Is it safe to let an AI agent use my logged-in email, bank or social media accounts?
  4. What can go wrong when an agent acts on my behalf, and how do I keep it safe?

Read with it

  1. Survey finding that 24% of US AI users use an agent regularly and 32% have let AI act for them without final approval, which is... menlovc.com Article 2026: The State of Consumer AI Menlo Ventures
  2. The single most important security idea for anyone deploying agents. simonwillison.net Article The lethal trifecta for AI agents Simon Willison
  3. Anthropic on letting Claude act in the browser without approving each step, with a safety classifier checking every action. claude.com Article Claude in Chrome is generally available Anthropic
  4. Mollick's way to decide when to delegate (your time against the time to brief and check) and the six questions a good brief... oneusefulthing.org Article Management as AI superpower Ethan Mollick (One Useful Thing)
  5. Google's guide to tasks in Gemini: who can use them, how its local and remote browsers work, and how to take over. support.google.com Guide Manage your tasks & workflows in Gemini Apps Google Gemini Help
  6. The same argument in shareable form for your team channel. simonw.substack.com Article The lethal trifecta (newsletter edition) Simon Willison
eChai Partner Brands The partnership →