Analysis
OpenAI rolled out ChatGPT Work, a cloud-based AI agent built on its GPT-5.6 model family that connects through MCP-based plugins to Gmail, Google Calendar, Slack and GitHub, giving the agent write access across a user's email, calendar, messaging and code repositories to complete complex, multi-step tasks independently. The product takes a stated outcome, breaks it into smaller steps, and can stay with a project for hours, producing finished documents, spreadsheets, presentations, reports and websites without ongoing human prompting.
The rollout begins with Pro, Enterprise and Edu users before expanding to Plus and Business tiers within days, and builds directly on OpenAI's earlier Workspace Agents product launched in April -- meaning ChatGPT Work represents an escalation in ambition rather than OpenAI's first attempt at enterprise agentic automation.
“Rather than continue maintaining a standalone product, OpenAI is redistributing Atlas's agentic browsing features into the ChatGPT desktop app and a new Chrome extension.”
In the same announcement window, OpenAI confirmed it is shutting down its standalone Atlas browser, which will stop working August 9, less than a year after its October 2025 launch. Atlas arrived with ambitions to bolt ChatGPT directly onto the web -- reading pages, rewriting them, eventually clicking through tasks itself -- but faced immediate security scrutiny, including prompt-injection attacks demonstrated within days of launch and a separate flaw that let malformed URLs expose users' browsing history. Rather than continue maintaining a standalone product, OpenAI is redistributing Atlas's agentic browsing features into the ChatGPT desktop app and a new Chrome extension.
The consolidation reflects cost discipline that predates this specific announcement: former Applications CEO Fidji Simo had directed the team to cut back on "side quests" months earlier, a mandate that also led OpenAI to shut down its Sora video-generation tool. For enterprise buyers, folding ChatGPT, Codex and Atlas's browsing capability into a single desktop app simplifies procurement and reduces the number of distinct OpenAI products IT teams need to separately evaluate and secure.
The bear case: bundling this much write access -- email, calendars, code, browsing -- into one agent significantly raises the blast radius of any single security failure, and the same GPT-5.6 model family powering ChatGPT Work is the one the UK's AI Security Institute found harboring universal cyber jailbreaks days earlier. What to watch next: how enterprises evaluating ChatGPT Work respond to that jailbreak finding before granting the agent production write access, and whether OpenAI publishes a security audit specific to ChatGPT Work's connected-app permissions model.
