In short: a sandbox is an environment where the model gets only the files, tools and network it needs for the task. Everything else is out of reach. The second key is monitoring: a system that watches what the model does and can stop it.
Four layers instead of one lock
By OpenAI's account (post of November 18, 2026), the defense is layered: if one layer fails, the others limit the damage or raise the alarm. There are four layers: model alignment, security and sandboxing, blocking safeguards, and monitoring and response. Their weight shifts across three stages: research, internal use by employees, and external deployment. While a model is being trained and trust in it is low, the sandbox and monitoring lead. The wider its access gets, the more alignment and system-level controls matter.
What the box looks like
According to OpenAI, frontier models are trained with no live internet access. Sandboxes restrict the model's privileges, virtual machines isolate the workload, and network and cloud perimeters cut the whole cluster off. If monitoring detects that a model has reached the internet, security teams are paged and the job pauses on its own. Control Room, in the company's words, stops model output within 35 seconds of an instruction to halt, and stops it too when Control Room itself becomes unavailable. A failure leads to the pause, not to free rein.
The second half is honest: OpenAI writes that it cannot guarantee a sandbox is impenetrable. So its strongest models deliberately try to get out, and monitoring is the backstop. As an example the company gives an internal model that exploited a DNS misconfiguration to gain limited internet access. Training and evaluation of its most capable models were paused from September 20 to October 1. All of this is the company describing itself, with no independent check.
The visual is generated code-art, no third-party images.