AI Agent Permissions: A Brief Is Not a Sandbox
Hi there.
AI agent permissions became much easier for me to think about once I separated instructions from access.
I can tell an assistant to work on one article and leave everything else alone, but that sentence is not a locked door around my files.
Here is how I now describe the boundary before a task starts, using a real publishing workflow rather than a promise that a separate setup is automatically isolated.
A separate setup is organization, not a locked room
My project instructions name an assigned task, an output folder, and files that must stay untouched.
They also prohibit opening secret files or changing unrelated schedules.
That gives the assistant a clear brief: what to do and where to stop.
A different configuration folder can keep settings organized.
It does not, by itself, prove that the process cannot reach other files on the computer.
I treat those as separate questions instead of describing every organized setup as a sandbox.
The distinction is practical.
A rule that says “do not open this file” tells the assistant how to behave.
An access boundary determines whether a particular operation can reach the file at all.
Both deserve attention, but they are not interchangeable.
I separate three layers before work begins
| Layer | Question I ask | What it establishes |
|---|---|---|
| Task brief | What work did I actually assign? | Intended scope |
| Tool permissions | Which operations does this tool allow? | Controls for those operations |
| Operating-system boundary | Which resources can this process reach? | Isolation within its configured coverage |
Claude Code's official sandbox documentation describes an operating-system-enforced boundary around shell commands and their child processes.
It also states that built-in file tools, MCP servers, and hooks run outside that shell sandbox.
The documentation distinguishes native Windows, where commands run unsandboxed, from running Claude Code inside WSL2 to use the sandbox.
I checked those statements on October 8, 2026.
Photo: Matheus Bertelli on Pexels. Historical interface shown for context, not evidence of current permissions.
For me, the useful question is not “does this product have a sandbox?”
It is “does the boundary cover the operation I am about to use?”
A shell restriction is not a reason to assume every file reader or helper process has the same restriction.
I inspect helpers before trusting their names
Before using an image helper, I checked its implementation rather than relying on its name.
It launched a separate official image-specialist process.
I kept its permission limited to bitmap generation, with research, writing, review, and publishing still handled in the main workflow.
That is the habit I want from an assistant: check what an available helper actually does before treating it as an allowed action.
A familiar name does not establish who will execute the work, which files will be passed along, or whether the task permits it.
For a non-developer, asking for that explanation is more useful than memorizing the command's spelling.
My task brief names both the work and the exclusions
I keep the brief short enough to read before every action.
The fields below are a reusable example, not a claim that writing them creates an operating-system security boundary.
- Assigned work: research and draft this article only.
- Allowed outputs: the article's task folder and its asset folder.
- Preserve: unrelated edits, other tasks, existing publication schedules, and existing settings.
- Do not open: passwords, keys, login files, or other secret material.
- Do not substitute: another executor or service when a required tool is unavailable.
- External actions: publish only after the exact final package passes the existing approval and review checks.
Photo: Matheus Bertelli on Pexels. Historical interface shown for context, not evidence of current permissions.
I also ask the assistant to distinguish preparation from publication.
A finished text file is a local artifact.
A public post changes an outside system, so it needs its own authorized step and a check of the actual result.
That keeps a useful draft from being described as a successfully published article.
A blocked step should stay visibly blocked
If a required tool is missing or its actual execution path falls outside the brief, I want a precise handoff.
It should say which artifact exists, which operation stopped, and what condition would allow that operation to resume.
“Use whatever works” is not a substitute for that record.
My current workflow therefore treats scope instructions as instructions, tool restrictions as tool restrictions, and isolation as something to verify within its actual coverage.
I do not turn a separate folder into a security claim.
And I do not turn permission to prepare an article into permission to change every connected account.
Which boundary do you make explicit first when you give an AI assistant access to a project?
I would love to hear the one rule that made your workflow easier to review.
Related reading:
Thanks for reading.
Comments
Post a Comment