Sources checked
Restrict files, network, and credentials
For an original repository workflow, start with an isolated checkout and only the files required for the task. Give network access according to the work that must happen, rather than exposing every credential available on the machine.
Permitted changes still need review
A successful sandboxed run can still produce a bad change inside the permitted area. Review the diff and run the relevant checks before promoting it.
What to fence
An agent that writes code will eventually write something you did not intend. Give it a filesystem that is not your laptop’s home directory, a network policy that is not “anywhere,” and credentials that are not your production keys.
Containers, separate OS users, and short-lived tokens are ordinary controls. They work only if the agent process cannot escape them.
Read-only is a meaningful mode. An agent that is supposed to explain a repository does not need permission to push.
An agent that drafts a query does not need permission to commit it. Start from no permissions and add the ones the task requires.
A default of “the developer’s own shell” is how a demo becomes an incident.
What a sandbox does not fix
A sandbox does not make the answer true. It limits damage.
Data you copy into the sandbox is data you have now exposed to the model and to the tool logs. Do not copy a production database in because the sandbox feels safe.
Copy the rows the task needs, or a synthetic set.
Network allow-lists go stale. A model that can fetch any URL can be steered by a page it fetched.
If browsing is required, restrict the hosts and treat the pages as untrusted input. Re-check the sandbox whenever you add a tool.
A new tool is a new door, even if the container name did not change.
What to record before you rely on it
Draw the sandbox: filesystem, hosts, secrets, and the user it runs as. Include a test that tries to read a file outside the fence and a test that tries to reach a host that is not allowed.
Record the pass. When you add a tool, update the drawing the same day.
Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.
Do not relax the check until the model happens to pass. Name a person who owns that failure.
A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.
Common questions
Is a virtual machine required?
Not always. You need a boundary that holds.
A separate account with no production credentials can be enough for a low-risk task. A code-executing agent that sees untrusted input needs a stronger fence.
Match the fence to the worst call, not to a product category.
Does hiding secrets in the prompt count as sandboxing?
No. The model and the logs can still reveal them, and the process can still use the ambient credentials of the machine. A secret the process cannot access is the control.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





