Understanding ChatGPT Work

OpenAI announced ChatGPT Work on July 9 and has been rapidly iterating since. The product is powerful yet extremely confusing, and the author spent considerable effort figuring out what it actually is.

Two products in one

ChatGPT Work comes in two flavors. The cloud version (Work Cloud) runs through the web or mobile apps; the local version (Work Local) requires the desktop app and can operate on your computer's files. The author focuses on the cloud version.

How does Work differ from Chat?

OpenAI says Chat is for Q&A and brainstorming, while Work handles tasks with a clear outcome. That explanation is nearly useless—the author has used Chat for those tasks for years. The real difference: Work offers a set of unique features. It can run code with internet access, launch a full Chrome browser, keep files across sessions, publish websites, run sub-agents, and schedule automated prompts.

How powerful are the new features?

The most exciting for the author is networked code execution: you can clone GitHub repos, install dependencies, and interact with the web. There is also a headless browser that can open sites, fill forms, take screenshots, and even run JavaScript against the page. Combined with a persistent filesystem and publishable ChatGPT Sites, this toolkit is quite capable.

Security concerns

Work combines access to private data, exposure to untrusted content, and an exfiltration channel—a perfect recipe for prompt injection attacks. OpenAI hasn't detailed its defenses, only mentioning an auto-review mechanism similar to Codex's.

Poor docs leave users guessing

The author complains that OpenAI explains Work's purpose but not its mechanics, and hides system prompts and tool descriptions. He resorted to having Work generate a site listing all 223 tools and 44 skills to understand it himself.

In short: ChatGPT Work is a powerful Swiss army knife, but you have to figure out how to use it before you can.