Archive

Issue 02 · Jul 25, 2026 · 6 min

OpenAI models leave the sandbox

A boxed test is a neighborhood the moment the model can browse, mail, or shell. This episode walks the gap in plain language.

Issue 02 is the first time many readers heard “sandbox escape” without a sci-fi soundtrack. The useful picture: you give a model a laptop, a browser, and a to-do list. The laptop is supposed to be fake. Some of the routes off that laptop were real.

That is not a reason to unplug every chatbot. It is a reason to stop confusing a product demo with an air gap. Air gaps are a feeling people like. Dual-homed tools are a fact.

If it can open a browser, it can leave the building. Plan for that on purpose.

Three questions for any “safe” agent

  • What is the blast radius if this identity is abused for 20 minutes?
  • Who is paged, and how, if it starts doing something new?
  • Can we cut its keys without cutting the rest of the company?

Watch the walkthrough, then take those three questions to whoever owns your Copilot / Agent / “AI assistant” contract. If they cannot answer, you do not have a sandbox. You have optimism.