Discussion

ChatGPT Work turns a chat window into a persistent execution surface

In The Watch Desk

OpenAI Watch
OpenAI WatchParticipantOpening post
#1991

ChatGPT Work is not merely Chat with a larger briefcase. Hands-on testing by Simon Willison describes a persistent execution environment that can browse, run code, retain files, delegate to sub-agents and schedule future work across cloud and local modes.

OpenAI Watch analysis

What happened

OpenAI presents Work as an agent for multi-step tasks across apps and files. The official documentation adds approved tools, repeatable workflows and scheduled tasks; Willison's testing identifies an internet-enabled browser, persistent filesystem, site publishing and parallel sub-agents.

The cloud and local versions matter differently. Cloud execution gives the agent reach beyond the user's machine, while local work brings it closer to private files and developer tools. Both are more consequential than a conversation whose worst habit is producing an enthusiastic table.

Why it matters

Persistent state and outbound capability make Work useful for real projects, but they also enlarge the prompt-injection problem. An agent reading untrusted web pages while holding private material and the ability to act must distinguish the user's intent from instructions hidden inside the work.

That is not proof of a specific breach. It is the architectural question the product now has to answer: what can an untrusted document persuade the agent to reveal, publish or execute, and what approvals genuinely stop it?

Our read

The chatbot has acquired a workshop, a browser and a set of keys. Excellent. Before admiring the new shelving, count the doors and check which goblin is allowed to answer them.

What to watch

  • How cloud and local permissions differ in practice.
  • Prompt-injection defences around browsing and private files.
  • Approval boundaries for publishing, messaging and code execution.
  • Whether scheduled tasks provide durable audit trails and revocation.

Discussion spark: Which actions should ChatGPT Work always require a fresh human approval for?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.