Cursor YOLO mode: what each run mode skips and how to limit auto-run
Cursor YOLO mode means the agent runs commands without asking you. Cursor's run modes today are Auto-review, Allowlist and Run Everything, and only Run Everything gives zero prompts. On your own Mac, pick Auto-review. Run modes are best-effort guardrails, not a hard security boundary, so commit your work before long runs and add a hook for commands you never want.
01The three run modes
You pick the mode in the desktop app under Settings > Agents > Approvals & Execution. In July 2025 the auto-run settings were an allowlist, a denylist and a checkbox against file deletion. Those have since been replaced by run modes. In Cursor 3.5 on May 22, 2026, Ask Every Time was deprecated and Run in Sandbox was folded into Allowlist. Cursor 3.6 shipped Auto-review as the recommended default a week later.
| Run mode | What runs without asking | Who decides the rest |
|---|---|---|
| Auto-review | Known-safe calls run, and shell commands are sandboxed when possible | A classifier reviews anything else |
| Allowlist | Only commands on your list, matched by the start of the command | An empty list behaves like the old Ask Every Time mode |
| Run Everything | Every tool call runs automatically | You. The table says you accept the risk and want zero prompts |
If you want no prompts at all, as in the old YOLO setup, pick Run Everything. Allowlist matching is case-sensitive: git matches git status but not gitk.
Outside run modes, Cursor stays cautious. Terminal commands need your approval by default. Every MCP connection needs approval too, and after that each tool call is approved one by one. An MCP server is a helper program that gives the agent extra tools.
02How Auto-review decides
Auto-review passes anything it cannot call safe to a classifier, which is a small model that judges whether an action should run. The classifier runs on Cursor's backend. It can allow a call you would have blocked, or block a call you would have allowed. So Auto-review cuts the number of prompts. It does not promise that nothing bad gets through.
03What the sandbox limits on a Mac
Auto-review and Allowlist can run shell commands inside a sandbox, a fence the operating system puts around a program. On macOS, Cursor builds that fence with Seatbelt, a macOS tool for restricting what a program may touch, started through sandbox-exec. A sandboxed command can work in your project. It cannot freely read protected files, write outside approved paths or contact arbitrary network destinations.
Network access is blocked by default and then opened by your network mode and a sandbox.json file. The default mode adds Cursor's built-in list for common package managers and language tools. That list includes ghcr.io, github.com, gitlab.com and golang.org. Allow All mode lets the sandbox reach everything.
Two limits matter:
- Reads are wide open by default. The agent can read files outside the workspace without approval.
- The fence has a door. Some commands need full access and bypass the sandbox; Cursor tells you and asks for approval. A sandboxed command that fails on a restriction can be rerun outside the sandbox, and the classifier reviews that rerun.
04Cloud Agents are different
Cursor Cloud Agents do not use run modes. A Cloud Agent runs inside its own dedicated machine, so it never asks you to approve an action. So the choice between Auto-review and Run Everything does not apply to it, and the questions on this page about your Mac's files do not either.
05Three protections to switch on
Cursor lists three switches that stop automatic actions:
- File-Deletion Protection stops the agent from automatically deleting files, including
rmcommands. - External-File Protection stops automatic changes outside the workspace.
- Browser Protection stops the agent from automatically running Browser tools.
They are extra guardrails, not guarantees. In a 2025 forum report, a user said delete protection did not stop a PowerShell Remove-Item call, and that the loss could not be recovered from a checkpoint.
06What went wrong in auto-run before
The Cursor YOLO mode stories that circulate are worth reading once, because they show the failure at its simplest. A user on the Cursor forum wrote that the agent tried to delete some old files, failed once, then ended up deleting everything on the computer, including itself. Cursor staff asked whether file deletion protection was on and recommended running YOLO in a virtual environment. In July 2025, Backslash Security found four ways for a compromised agent to get around Cursor's old denylist and run unauthorized commands.
Take one lesson from this. Prompt injection is when text the agent reads, such as a web page or a README, carries instructions the agent then follows as if you had typed them. A list of forbidden commands and a delete checkbox are good guards against honest mistakes. They are not a wall against an agent steered that way, so keep Git history and isolate anything you cannot afford to lose.
07How to limit Cursor YOLO mode
If you want fewer prompts without giving up control:
- Start with Auto-review. It is the recommended default.
- Commit before long runs. Agents can modify workspace files without approval, except configuration files, and changes save immediately to disk. Cursor says to use version control so you can revert.
- Check auto-reload. If it is on, agent changes might execute before you review them.
- Add a hook for the commands you never want. A hook is a small script that Cursor runs before an action and that can stop it. Cursor's
beforeShellExecutionhook controls shell commands. The step-by-step mechanics are on the Claude Code hooks page, and Cursor's version works the same way in spirit. - Combine soft and hard controls. Pair Auto-review and allowlists with approvals, hooks and sandboxing instead of relying on one layer.
- Keep untrusted repos out. Workspace trust is disabled by default. For untrusted repos, use a basic text editor instead.
08Which mode for which place
Cursor YOLO mode, meaning Run Everything, fits a disposable virtual machine with no keys in it, the setup Cursor staff pointed to. It does not fit the Mac where your SSH keys, cloud credentials and browser sessions live. On that Mac, stay on Auto-review.
Other agents have their own version of the same switch. In Claude Code it is the bypass flag, and Codex has its own, covered in Codex full auto. Cursor itself has had sandbox escape bugs, including one fixed in version 2.5, so keep it updated. The advisory list lives on the security page. For isolation in Cursor, see Cursor sandbox and is Cursor safe. The wider settings picture is in AI agent permissions and Claude Code permissions.
09Should you leave Cursor YOLO mode on?
No, not on a Mac you care about. Use Run Everything only inside a throwaway environment, and on your main Mac stay on Auto-review. Before any long unattended run, commit your work, switch on the three protections above and make sure the agent cannot reach a folder you would miss.
With Run Everything on, it helps to know what actually ran on the Mac. Artificial Immunity's free, open-source core watches Cursor, Claude Code, Codex, MCP servers and every program start, and reports what it finds without blocking anything.
FAQ
What is Yolo mode?
Yolo mode is the common name for letting an AI agent run terminal commands without asking you first. In Cursor, the run mode with zero prompts is Run Everything. Cursor's table says it is for people who accept the risk.
What are the different modes in Cursor?
There are three run modes: Auto-review, Allowlist and Run Everything. You set them in Settings > Agents > Approvals & Execution.
How do I enable YOLO mode in Claude Code?
Start Claude Code with the --dangerously-skip-permissions flag, which is the same as --permission-mode bypassPermissions. Use it only in a container, VM or dev container without internet access. Even then, actions that no mode auto-approves still apply. The full story is on the dangerously-skip-permissions page.
Which Cursor run mode is the safest?
Auto-review. Cursor calls it the safest useful setup for most people, though it is not a security boundary and its classifier can make mistakes.
Do run modes apply to Cursor Cloud Agents?
No. A Cloud Agent runs inside its own dedicated machine, so it never asks you to approve an action.
Related guides
Sources
- Cursor docs: run modes
- Cursor docs: agent security
- Cursor docs: permissions reference
- Cursor docs: hooks
- Cursor docs: enterprise security hardening
- Cursor docs: CLI changelog
- Claude Code docs: permission modes
- Cursor forum: deleted everything thread
- Cursor forum: delete protection thread
- The Register: Cursor safeguards bypassed
- Artificial Immunity