Analysis · Hacker News ·

ChatGPT Work: internet access, headless Chrome, persistent files

Simon Willison's deep dive reveals ChatGPT Work's hidden superpowers: full internet access, headless Chrome, persistent filesystems, and 223 registered tools — plus serious safety questions.

Based on reporting by Hacker News — analysis by dalili

OpenAI's ChatGPT Work, launched July 9th and available to $20/month+ subscribers, is far more powerful than most users realize. Developer Simon Willison has published the most thorough breakdown yet, revealing features that transform it from a chatbot into a full development environment.

The biggest revelation: ChatGPT Work's code execution environment has open internet access by default. Unlike ChatGPT Chat, which blocks external API calls, Work can clone GitHub repos, install packages, and interact with any website. It also runs a full headless Chrome browser that can fill forms, take screenshots, and execute JavaScript against live DOMs.

Work sessions get persistent filesystems shared between sessions — Willison counts 171 scratch folders in his workspace. It can build and deploy entire websites via Cloudflare Workers, run sub-agent sessions with different models (Sol, Luna, Terra), and schedule recurring prompt automations.

Willison flags serious safety concerns: Work combines access to private data, exposure to untrusted content, and communication channels — his "lethal trifecta" of agent risk. He notes OpenAI hasn't explained how it protects against prompt injection attacks in this environment.

Key takeaways

  • Work's code environment has open internet access — unlike ChatGPT Chat
  • Headless Chrome can fill forms, run JS, and take screenshots
  • Persistent filesystem shared across sessions (171+ folders)
  • 223 registered tools and 44 skills discovered in the system

Why it matters

ChatGPT Work quietly crosses a threshold: it's the first mainstream AI product that gives users a general-purpose agent with internet access and code execution. The safety implications are significant — Willison's lethal trifecta framework highlights real prompt injection risks that OpenAI hasn't publicly addressed.