Tools & responsibility
Check which files an AI assistant can read before you give it work
OpenAI’s new explanation of the Codex sandbox shows why permission settings matter. A limit on editing files does not necessarily limit reading them.

Before asking an AI assistant to organise business files, establish what it can read, change and send outside the computer. A request to touch only one folder is useful direction; the permission settings must enforce the intended limit.
In its 18 November explanation of the Codex sandbox, OpenAI describes controls over file access and network connections. Author Michael Bolin describes “limits you control”. A sandbox restricts what a running program can access. The company explains that one built-in workspace profile, a set of access settings, allows changes within the working folder but broader reading elsewhere, subject to the computer user’s existing permissions. Network access is disabled in that profile.
Read access matters as much as write access
These are documented settings for Codex commands, instructions executed by the computer, not a description of every AI product or every installation. Check the actual profile your technical team has selected. OpenAI also says harmful actions remain possible inside the permitted area. A sandbox reduces access; it does not establish that the work is correct.
Connected external tools need their own permission review. A restriction on local commands does not by itself establish the limits of a separate service account, an account used by software. For a small firm, the immediate consequence is practical: do not use a folder containing client material merely because the assistant has been told to ignore it.
Use fictional invoices for the first trial
Ask your technical lead to prepare a test workspace with two text files: invoice-a.txt contains “Fictional invoice A: USD 12”; invoice-b.txt contains “Fictional invoice B: USD 18”. Use no customer records. The lead should configure and verify the permitted reading area, output folder and network restrictions before the trial. If those limits cannot be established, use an ordinary chat with the fictional text instead of granting file access.
Then use this prompt: “Read the two fictional invoice files. Create total.csv in the approved output folder with the header currency,total and one data row giving the currency and calculated sum. Do not change either input file. Report any missing or conflicting information instead of guessing.”
Open the result in a text editor and compare it with the supplied values: 12 + 18 = 30. Check that the original files are unchanged. This checks the requested output and preservation of inputs. It is not a security test or proof that other files were inaccessible; the technical lead must verify permissions separately.
Find AI use outside the approved setup
The Shadow AI Risk Diagnostic helps managers examine unapproved AI use. Prepare a list of approved tools, the kinds of information staff handle and the current access rules. Answer its questions without entering personal or confidential records, then use the section results to choose a control to investigate.
The result provides a score and an exposure category with recommendations. These reflect your answers; the tool does not detect software or audit a computer. A sensible next step is to ask the system owner to compare the approved access list with actual accounts and remove unnecessary access through the organisation’s normal process.
Sources and further reading
Sources checked 19 November 2026.