OpenAI Named Its Agent After What It Cannot Do
There is a phrase most companies bury deep in the fine print. OpenAI put it in the headline.
When OpenAI launched its flagship coding agent in April 2026, it named it “Codex for (almost) everything.” One word. Four letters. The most expensive confession in software product naming since Microsoft called its search engine Bing.
OpenAI has spent two years building the market case for autonomous AI. Autonomous agents. Autonomous pipelines. Autonomous software development. The company that shipped the product category is also the company that named its primary product after the category it cannot reach.
The “almost” is not marketing modesty. Read the official documentation and you find the technical admission underneath it. Codex runs tasks inside isolated, sandboxed containers. During the agent execution phase, internet access is disabled by default. The agent that is supposed to autonomously build your product cannot, by default, reach the APIs your product talks to. It codes against what you gave it at setup. The live world is locked out while it works.
This is not a bug. It is the architecture. OpenAI built it this way because giving a coding agent unrestricted internet access during execution creates a threat surface no enterprise security team would approve. So the product that was supposed to end the human-in-the-loop ships with the internet turned off while the agent runs.
By May 14, 2026, Codex was live on mobile for every ChatGPT user. 800 million weekly active users now have access to the autonomous coding agent. The one that cannot reach the internet while it works unless you go into settings and unlock it per environment.
The companies whose entire thesis was “agents will access live systems autonomously, at scale, without human checkpoints” are building on a product whose official title contains the word that ends that thesis.
“Almost” was always the answer. OpenAI just said it out loud.
👉 Read the agentic security news. Instantly analyze the threat vector, see if it applies to your setup, and find the gaps with our interactive playbook. All free.
👉 Follow me on LinkedIn | X | Substack for weekly analysis of real agent failures, control gaps, and what the frameworks are and are not catching.