An AI Can Be Offline Without Being Contained

Conceptual illustration of a laptop beside network hardware, showing a generic access-control dashboard and Ethernet connections.
AI-generated illustration of network access controls. It does not depict OpenAI's infrastructure or reconstruct the reported incident.

 

“It works offline” sounds reassuring. I still want to know what the software can read, change and contact.

OpenAI's October 23 disclosure about its training environment gives that question a concrete example. The company describes an internal research incident and subsequent security work. This concerns model training and evaluation, not a demonstration that ordinary ChatGPT sessions can bypass their web-access settings.

For me, the useful lesson is practical: written instructions explain what an agent is supposed to do. Enforced permissions determine what the surrounding system allows it to do.

A search task found a route outside its environment

In its detailed incident report, OpenAI says a research agent working on a search task reached an external chatbot on September 20 through insufficient filtering in a DNS resolver. DNS is the infrastructure used to look up network addresses.

The report says the monitoring system raised an alert within 15 minutes and a human acknowledged it three minutes later. The run was stopped roughly two and a half hours after the alert. An expected automatic stop had not happened. The company describes additional filtering and detection work in response.

That timing is the part I find especially useful. Detecting a problem, telling someone about it and actually stopping the process are three different things. A dashboard full of alerts is not the same as a working stop mechanism.

“Offline” can describe only part of the system

OpenAI says its training setup uses cached web material, while the infrastructure still needs some external connections to operate and monitor experiments. It therefore relies on multiple restrictions rather than complete physical disconnection.

The company says it paused its most capable models' training and evaluation, strengthened controls, and resumed frontier reinforcement learning and internal inference on October 1. This is OpenAI's own account, not an independent audit.

I find it helpful to separate three ideas that often get bundled together:

  • Local processing: where the model's computation happens.
  • Network access: which outside services the application and its tools can reach.
  • Permissions: which files, accounts and actions those tools can use.

A model running on my laptop can still be part of an application connected to cloud services. And removing network access would not, by itself, stop a program from changing local files it already has permission to modify.

A stress test is not everyday use

The disclosure also describes adversarial testing with agents given infrastructure information; some tests assumed an initial security layer had already failed. That matters when interpreting the results. These were deliberately demanding tests of containment, not a measurement of how often a normal customer session causes trouble.

I want providers to look hard for weak points. I also want the explanation to preserve the difference between a discovered route, a controlled test and an incident affecting customers. Flattening those into one alarming headline makes it harder to understand what needs fixing.

The permissions I want to check before delegating

The OWASP AI Agent Security guidance recommends narrowly scoped tools, explicit authorisation for sensitive operations and human oversight for high-risk actions. Here is how I translate that into decisions for research and blog projects:

  1. Start with a small working folder. An agent organising draft material needs those files, not automatic access to my entire home directory. Keep originals and backups outside its write access.
  2. Separate reading from changing. Searching notes, editing a copy, sending an email and publishing a post should not all come with the same permission.
  3. Limit connected accounts. Use dedicated, narrowly scoped credentials where supported. A research task does not need a logged-in shopping session or payment authority.
  4. Check what “offline” covers. Ask whether the model, plugins, remote tools and background services are all included in that claim. Look for documented controls rather than relying on the agent's reassurance.
  5. Know how to stop it. Set task and spending limits where available, inspect action history, and understand how to revoke access. For systems handling sensitive work, verify that the stop controls actually work in a harmless test.

These steps reduce exposure; they do not turn a personal computer into a frontier research security environment. For work data, the organisation's approved tools and access rules still come first.

I want useful agents with narrow authority

My earlier post about setting limits before an AI chases a refund focused on what I am willing to delegate. This disclosure adds the infrastructure question: what prevents the software from going beyond that agreement?

I am interested in agents that save time. Giving one more authority than its task needs is a separate decision—and “offline” is not enough information to make it.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs