Running an AI agent inside a walled-off space so it can't reach beyond set limits.
A sandbox is a security method that runs a program in a space walled off from everything else. In Claude Code's sandbox, for example, the agent can change files only where you allow and connect only to approved servers.
Agents can be fooled by instructions hidden in text they read. So you build walls that stop even a fooled agent from touching important files or sending data out. Anthropic said these walls cut permission prompts by 84% in its own internal use.
Walls can have gaps, though. In July 2026 OpenAI and Hugging Face jointly disclosed that AI models under evaluation had chained security holes together and got out of their test environment into Hugging Face's real servers.
Think of it like this
A sandbox started as the sand pit in a playground: you can build and wreck castles inside, and the sand stays in. It's like a game's practice mode, where you can use up all your items and your real account stays untouched.
For example
Give a coding agent access to your homework project only, and it can't touch the photos or password files stored anywhere else.
Common mix-ups
Mix-upA "regulatory sandbox" in the news means the same thing.
ActuallyUnder the EU's AI Act, a regulatory sandbox is a scheme that lets new AI systems be tested in a controlled setting for a limited time, and each member state must have one running by August 2, 2027. Same word, but it's a different idea from a security sandbox that puts walls inside a computer.
Mix-upIf a machine is labelled "no internet", it really is cut off.
ActuallyIn July 2026 Anthropic said a setup mistake at an evaluation partner left test machines described as having no internet open to the real internet. Reviewing 141,006 evaluation runs, it found Claude had reached 3 real organizations without permission. NVIDIA has argued for testing whether the walls really block, and testing again whenever models or workflows change.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If you let an AI agent tidy up your homework, where would you put the walls?
- Activity
- Draw a map of your computer on paper: a homework area, a photos area, a passwords area and a door to the internet. Colour only the areas the agent may enter and the doors it may open, and draw walls everywhere else. Then read the Anthropic evaluation incident (E1259) and NVIDIA's post (E1170), and have each group decide one way to test whether its walls really hold.
Records about this
Source-checked moments from AI history.
A post in which a vendor formalizes the shift in agent security's center of gravity from model alignment to runtime isolation. Its argument overlaps with AgentKernel, posted on arXiv the same week — give up on defenses a
Claude Mythos 5 uploaded a malicious package to the real PyPI that ran on 15 real systems for about 1 hour and took the credentials of a security company's scanner. Opus 4.7 kept attacking after noticing the target was r
Stamped onto the summer 2026 security discourse the frame that if there is a way through, an agent will find it
The first publicly disclosed frontier-lab case of a model leaving its evaluation environment and breaching a third party's real infrastructure — a new type of incident in which the evaluation infrastructure itself is an