Michael JohnsonYou can start an agent task on your computer, continue it on the web or a mobile device, while your computer sleeps. That is how OpenAI describes Codex Cloud in its September 29, 2026 release notes. This is more than a change of interface: the agent’s work is no longer tied to an open local session.
For developers, this means more than convenience. When execution is separated from the device, it becomes more important to understand where it happens, what it can access, and how to verify the result. The published description confirms some properties of Codex Cloud, but does not allow its security or reliability to be assessed independently.
What OpenAI announced
According to OpenAI’s release notes, Codex Cloud lets users start and continue tasks from a computer, the web, or a mobile device. Projects can use reusable environments with repositories, tools, and dependencies; each task gets its own isolated workspace. OpenAI also says that tasks can run while the user’s computer sleeps.
TechCrunch described the Codex update at DevDay as a move toward reusable, more persistent, customizable cloud environments. However, that report is no substitute for a technical test of isolation or an assessment of the agent’s behavior in a real project.
From a local session to a cloud task
Cloud execution lets you switch between devices without keeping your computer turned on. But it raises more questions between starting a task and checking its result: How can you monitor its progress? What permissions does the environment have? And how can you stop the work if something goes wrong?
These are not claims about shortcomings in Codex Cloud, but practical questions that become more important when a task continues without an open local session. For a team, it matters not only whether an agent can write code, but also how easily changes can be reviewed, access to repositories and tools controlled, and results reproduced or rolled back. The available materials do not describe these processes in detail.
“Isolated” does not mean “independently verified”
OpenAI describes each task’s workspace as isolated. This is a description of the product’s design, but on its own it neither reveals the threat model nor proves that the safeguards are effective. Assessing risks requires details about isolation boundaries, network and secret access, data handling, and recovery after a failure. The published description of Codex Cloud contains no such analysis.
So the justified conclusion is limited: OpenAI says tasks run in separate workspaces, but the available sources provide no independent assessment of the implementation. That is no reason to declare the feature unsafe—just as the word “isolated” alone is no reason to treat it as a sufficient guarantee.
A separate security context
On September 29, the Associated Press reported that the release of another OpenAI model had been delayed amid safety concerns. This is relevant context for discussing agentic systems, but it is not evidence of risks in Codex Cloud: the sources establish no causal link between the two events and do not show that they share the same risk profile.
Codex Cloud should be assessed based on documentation, verifiable testing, and user experience—not by transferring concerns about another product to it.
What to check before using it
When evaluating an agent’s cloud-based work, it is useful to ask:
- Which repositories, tools, and data can the task access?
- Where are the environment’s restrictions and isolation boundaries documented?
- How can changes be reviewed and checked before they are accepted?
- How can you stop a task and recover from unwanted changes?
These are questions for checking the documentation and workflow, not a list of confirmed Codex Cloud features. The stated shift is that an agent can continue a task while the computer sleeps, and that work is accessible across devices. For a team, autonomy is only one part of the assessment: observability, access control, and result verification matter just as much. The published information describes the product’s direction, but it is too early to draw a final conclusion about its reliability.