Least privilege means giving a personal AI agent only the access a task needs, for only as long as it needs it. In practice that is four habits: connect fewer apps, choose read-only scopes where you can, prefer one-time or time-bounded approvals over permanent ones, and take access away when you are done. It does not stop an agent from making mistakes or being tricked, but it limits how far the damage can reach. This guide applies the principle to Meta Muse, OpenAI dots, Claude Duet, Grok Bot and OpenClaw, based on official docs as of November 2026.
Why does least privilege matter more for agents than for apps?
A normal app does what its code says. An agent decides what to do based on what it reads, and what it reads can include hostile instructions hidden in an email, web page or document. That is indirect prompt injection, and no vendor claims to have eliminated it. OpenAI says dots defences "reduce" but "do not eliminate" the risk; xAI says Grok Bot safeguards "reduce, but do not eliminate, risk from malicious content".
If you cannot fully stop an agent from being misled, the next best thing is to limit what a misled agent can touch. OWASP lists this as "Excessive Agency" for language model apps (LLM06), and its 2025 agentic list includes "Identity & Privilege Abuse" (ASI03). See our OWASP agentic top 10 explainer for the full list.
The joint guidance Careful adoption of agentic AI services, published 1 May 2026 by ASD's ACSC, CISA, the NSA, the Canadian Cyber Centre, NCSC-NZ and NCSC-UK, puts it plainly: "strict adherence to the principle of least privilege is critical", and agents should never be granted "broad or unrestricted access, especially to sensitive data or critical systems". That guidance is aimed at organisations, not households, but the reasoning carries over directly.
- Agent reads hostile content
- Agent is misled
- Tries an action
- Narrow scopes block or limit it
The four dimensions of an agent's access
It helps to break "access" into four separate questions.
1. Which accounts can it reach?
Every connector (Gmail, Outlook, Drive, Slack, a bank-adjacent app) adds data the agent can read and actions it can take. Meta's advice for Muse is to connect apps one at a time, per its launch post. OpenAI's dots can reach thousands of apps through plugins, and plugin permissions are shared across ChatGPT, ChatGPT Work and Codex, so a connection made in one place may be live in the others.
Rule of thumb: if you cannot name a task this week that needs a connector, do not connect it.
2. What can it do in each account?
Within a connector, choose the narrowest scope. Muse lets you pick read versus send per connector; its Mac Messages connector is opt-in, has its own access levels including None, and needs Full Disk Access, which is a lot of reach for a single feature. Our guide to read-only vs send access goes through the email and calendar options.
3. How long does each permission last?
Meta's security post lists five approval scopes for Muse: one-time, session, task, time-bounded and perpetual, each tied to a connector or destination. Shorter is safer. A perpetual approval to message a destination is effectively an "Always allow" (see why "Always allow" is risky).
4. Where does the agent run?
An agent with its own cloud computer is contained to that computer. An agent with access to your own laptop is not. dots' local computer connection is optional and off by default. Grok Bot runs local desktop commands only if you allow them in Settings → Computer. OpenClaw runs on your machine, and in 2.0 its sandbox is off by default.
| Dimension | Narrow choice | Broad choice |
|---|---|---|
| Accounts | One connector for one task | Everything connected "just in case" |
| Scope | Read, or read plus draft | Read, send, delete |
| Duration | One-time or time-bounded | Perpetual or "Always allow" |
| Location | Vendor cloud computer, sandbox on | Your own computer, no sandbox |
Approval and scope options as documented by each vendor, as of November 2026.
How do I apply least privilege to each agent?
Meta Muse
Use per-connector access levels and pick read-only first. Prefer one-time, task or time-bounded approvals in the client UI rather than perpetual ones. Meta's Sentinel is the only component that can grant permission, and it removes auto-allow from processes that have read your data ("tainted egress"), which helps, but it acts on the permissions you have given. Details are in our Muse safety settings guide.
OpenAI dots
Keep Custom Rules narrow. Reviewers at eesel found that broad rules were rejected, which is a sign the system expects specific ones. Leave the local computer connection off unless a task needs it. See setting up dots Custom Rules safely.
Claude Duet
Connectors carry over from your Claude account, and switching an app off in Claude also switches it off in Duet. Review what is connected in Claude before you start. Duet's built-in limits are strict (no moving money between accounts, no trading securities), so the main lever is how many apps you connect. See Claude Duet approval settings.
Grok Bot
Keep Settings → Computer on "Ask every time". Be aware that all of a user's Bots share one cloud computer, including files, browser sessions and logins, and xAI says "Do not use separate Bots as a security boundary." If you want true separation for a sensitive task, the docs say shared or team Bots get their own computer. See Grok Bot computer access settings.
OpenClaw
Least privilege is your job. Turn the sandbox on, set the workspace to none or read-only, use an allowlist for commands instead of full mode, deny tools you do not need, and run the gateway as a low-privilege OS user. The OpenClaw sandbox and command approvals guide covers each setting.
- List tasksWrite down what you actually want the agent to do this month.
- Map accessFor each task, note the one connector and scope it needs.
- Disconnect the restRemove connectors with no task attached.
- Shorten approvalsReplace perpetual approvals with one-time or time-bounded ones.
- Review weeklyCheck the activity log and prune again.
Does least privilege break the lethal trifecta?
Partly. Simon Willison's lethal trifecta is private data plus untrusted content plus a way to send data out. If an agent has all three, a hidden instruction can make it leak your data. Least privilege lets you remove one leg for a given task:
- A research task that reads the web does not need your inbox connected.
- An inbox summary does not need send permission.
- A drafting task does not need web browsing.
Our lethal trifecta explainer has more examples.
What least privilege does not fix
- Memory. Disconnecting an app does not necessarily erase what the agent learned from it. OpenAI says disconnecting an app from a dot doesn't remove what it learned; reset or delete the dot instead. Anthropic says Duet keeps what it learned until you delete or reset. See deleting agent memory.
- Approved mistakes. If you approve a wrong action, scopes do not help. Read approval prompts carefully.
- Vendor access. Least privilege controls what the agent can do, not what the vendor can see. Meta's architecture, for instance, does not prevent Meta from accessing data.
Bottom line
Least privilege is the one safety habit that works across every agent, whatever its defences. Connect less, grant narrower, approve for shorter, and clean up afterwards. For which tasks to hand over in the first place, see what to let an AI agent do. For a vendor-by-vendor view of default permissions, see the AI agent safety and permissions guide on our sister site.



