OpenAI’s Operator system, launched in late 2025, could browse the web and fill out forms. Its successor, launched in April 2026 under the internal name Atlas and now available to enterprise customers, can control an entire computer for hours — writing and executing code, managing files, conducting research across dozens of sources, sending emails and scheduling meetings, and correcting its own errors when tasks do not go as planned.
The demos are remarkable. In one video shared by a beta customer, Atlas spent four hours preparing a market research report: it searched academic databases, pulled financial filings, read and summarised competitor websites, built financial models in Excel, created visualisations, and produced a 40-page document. A human analyst would have taken a week.
“The thing that surprised us was the error recovery,” said one enterprise customer who has been testing the system. “It doesn’t just blindly execute. When something doesn’t work, it stops, diagnoses what went wrong, and tries a different approach. That’s not how any previous automated system has behaved.”
The safety implications are significant. An AI system that can take actions on a computer — send emails, make purchases, modify files, execute code — can cause real harm if it misunderstands instructions or encounters unexpected situations. OpenAI has implemented several layers of safeguards: the system requires explicit human approval before certain categories of action, maintains a detailed log of every step it takes, and can be interrupted at any point.
Whether those safeguards are sufficient is the subject of active debate. Several prominent AI safety researchers have published critiques arguing that the approval mechanisms are too easy to bypass in practice and that the logging system, while comprehensive, is too complex for most users to meaningfully audit. OpenAI has responded that no autonomous system is risk-free and that the appropriate comparison is not to a perfect AI but to the humans performing the same tasks.
Leave a Reply