Get the newsletter
OpenAI · AI / ML · Filed Sep 10, 2025 · USPTO published Nov 19, 2026 · This week's release

OpenAI Patents a Locked-Down Workspace Where AI Writes Code for Human Approval

OpenAI's patent application describes an AI coder that works with no internet and no direct access to your real project, then hands you a list of changes to approve.

A coding screen presenting AI-suggested software edits and test results for a developer to inspect and approve. Drawing from patent filing US 2026/0367910 A1.
A coding screen presenting AI-suggested software edits and test results for a developer to inspect and approve.All 11 drawings ↓
Publication number
US 2026/0367910 A1(US20260367910A1)
Applicant
OpenAI OpCo, LLC
Filed · Published
Sep 10, 2025 · Nov 19, 2026
Inventors
Bryan Ashley, Channing Conger, Hantao Wang
Status when we published
Waiting for an examiner
Document
20 claims
US classification
717/101
Parent application
Claims priority from a provisional application 63806537 (filed May 15, 2025)

What OpenAI's sealed-workspace coding setup does

You ask an AI assistant to upgrade the software packages behind your app, then walk away for lunch. When you get back, the worry is obvious: what did it touch while you were gone, and could it have reached out to the internet or your real files?

This OpenAI patent application describes a setup where the AI works inside a sealed copy of your project. The copy gets everything it needs first, then its internet connection is switched off before the AI starts. When the AI says it's done, ordinary software (not AI) compares the before and after and produces a plain list of what changed.

You read that list along with a short summary. Only if you click a button does it become a proposal (what developers call a pull request) that your team can review before anything touches the real project.

From the filing · CLAIM 1
obtaining, from a machine learning model, an indication that a software development task provided to the machine learning model is completed, wherein the software development task comprises instructions in natural language requesting to modify code in a repository; …

Translation: The AI finishes a coding job that was requested using plain English instructions.

How the sealed container and patch step work

The system starts when someone gives the AI a task in everyday language, such as "upgrade this server's packages." It spins up a container (a sealed, disposable virtual computer) built from the user's chosen setup, installs the project's tools, and clones (copies) the code repository into it. Then it switches off the container's external network connection before the AI is handed the task. The AI keeps an internal channel, so it can still read files, edit code and run tests.

When the AI signals that it has finished, ordinary non-AI software (a "deterministic application," meaning one that gives the same answer every time) compares the container's files to the original and generates a diff, a line-by-line list of what was added, removed or changed. A screen shows that diff plus a summary for human review. The filing also describes the AI trying several variants of the same task so you can compare them.

If you approve, the system builds a change proposal made of:

  • A manifest listing the repository's files after the change
  • A commit that points to a snapshot (a "tree") and its parent, so the project's history stays intact
  • The diff itself

A second container with a fresh clone of the repository can then apply the proposal, run tests, and handle added files, removed files and binary files (images, compiled code). That container holds the login credentials, which the AI never sees.

From the filing · THE ABSTRACT
… generating a diff, using a deterministic application, illustrating changes generated by the machine learning model to the code in a cloned repository, wherein the diff comprises at least one change to a text file associated with the repository; …

Translation: A separate program checks the AI's work by creating a side by side comparison of the updated code.

Why this matters for teams letting AI edit code

Companies hesitate to let AI touch real software because one bad move could leak passwords, corrupt a project or push broken code. This design addresses that by keeping the AI away from the internet, your credentials and the live project, and by keeping a person in the approval seat. If you manage a team's code, that could mean reviewing one clear list of changes instead of trusting an AI with the keys.

It also fits how developers already work. The output arrives as a standard proposal that teammates can comment on, test and roll back. The filing frames this separation as meeting business security requirements, which shows how much of the effort in AI coding tools goes into trust and oversight rather than the writing of code itself.

This is the 11th OpenAI filing we've tracked since September in our AI guardrails race, following their code security checker and self-testing bug hunter.

Source: US 2026/0367910 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time OpenAI patent news breaks.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings · 11 sheets

11 drawing sheets from US 2026/0367910 A1 · click any drawing to enlarge

How user prompts clear safety checks before reaching the AI, drawing sheet 1 from US 2026/0367910 A1
Sheet 1 How user prompts clear safety checks before reaching the AI
Locking an AI coder in an offline workspace, drawing sheet 2 from US 2026/0367910 A1
Sheet 2 Locking an AI coder in an offline workspace
Safe steps for letting an AI write code, drawing sheet 3 from US 2026/0367910 A1
Sheet 3 Safe steps for letting an AI write code
Reviewing AI code edits and sending them for team approval, drawing sheet 4 from US 2026/0367910 A1
Sheet 4 Reviewing AI code edits and sending them for team approval
Asking the AI to generate multiple coding solutions to compare, drawing sheet 5 from US 2026/0367910 A1
Sheet 5 Asking the AI to generate multiple coding solutions to compare
Reviewing code in a separate area before updating projects, drawing sheet 6 from US 2026/0367910 A1
Sheet 6 Reviewing code in a separate area before updating projects
Turning plain-language requests into specific line-by-line file edits, drawing sheet 7 from US 2026/0367910 A1
Sheet 7 Turning plain-language requests into specific line-by-line file edits
Training the AI models behind automated coding tools, drawing sheet 8 from US 2026/0367910 A1
Sheet 8 Training the AI models behind automated coding tools
Show the other 3 sheets
How the AI model turns incoming text into code, drawing sheet 9 from US 2026/0367910 A1
Sheet 9 How the AI model turns incoming text into code
Inside the building blocks that process and generate software, drawing sheet 10 from US 2026/0367910 A1
Sheet 10 Inside the building blocks that process and generate software
Standard computer hardware that runs the coding service, drawing sheet 11 from US 2026/0367910 A1
Sheet 11 Standard computer hardware that runs the coding service
Patent filing page

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.