Is Cursor safe? Security record, settings and real risks
Is Cursor safe? For everyday work, yes, if you pick a safer run mode and turn on Privacy Mode, which stops Cursor training on your data. Cursor recommends the Auto-review mode. Its own docs call run modes best-effort guardrails, not a hard security boundary, so add the protections below and keep the network closed. The vendor holds independent security certifications and an audit attestation.
An editor that runs commands on its own is a lot to trust. This page shows what Cursor's agent can reach on your Mac, what the vendor publishes about security and data, and the incidents on record. It ends with five settings that lower the risk. Is Cursor safe enough for a work repository? Yes, with the limits below, and the sections follow the order in which you would meet them. If you use other agents too, the same question is answered for Claude Code, Codex and Copilot.
01What Cursor's agent can touch on your Mac
The agent works in your workspace and can run terminal commands. How much it does without asking depends on the run mode, set in Settings > Agents > Approvals & Execution. Since Cursor 3.6 (May 29 2026), Auto-review is the recommended default.
| Run mode | What it means |
|---|---|
| Auto-review | Known-safe calls run, shell commands go into the sandbox when they can, and anything else goes to a classifier, a second AI model, for review. It is the safest useful setup for most people. |
| Allowlist | Commands you have listed run without asking; anything else asks first. Entries use case-sensitive prefix matching, so git matches git status but not gitk. |
| Run Everything | Every tool call runs automatically. This is the "YOLO" setup from older forum threads, the one where the agent never asks. |
Cursor 3.5 deprecated Ask Every Time and folded Run in Sandbox into Allowlist. Cloud Agents, which run on Cursor's own machines instead of yours, do not use run modes.
Other defaults worth knowing:
- Agents can modify workspace files without approval, except config files. Use version control.
- MCP connections and each MCP tool call need approval by default. An MCP server is a program that gives the agent extra tools.
- Read access defaults to System, so the agent can read files outside the workspace without approval. Narrowing it to the workspace needs Cursor 3.23 or later.
- Workspace trust is disabled by default, and restricted mode breaks AI features.
- With auto-reload on, agent changes might execute before you review them.
02Is Cursor legitimate?
Yes, and the evidence is checkable. Cursor holds ISO/IEC 27001:2022 (the international standard for managing information security), ISO/IEC 42001:2023 (the international standard for how a company manages its AI) and AIUC-1 (a security certification for AI agents), plus a SOC 2 Type II attestation, an independent auditor's report on its security controls. It publishes security advisories on GitHub, including critical ones, and it published a security notice about a tampered marketplace extension. That open record is what lets you judge it.
Legitimate does not mean unbreakable. Auto-review is not a security boundary and the classifier can make mistakes, in both directions. The enterprise hardening guide tells you to pair these best-effort guardrails with deterministic ones.
03Is your data safe with Cursor?
Privacy Mode is the setting that answers this. It is available to anyone, free or Pro, and when it is enabled Cursor will not train on your data. Turn it on before you open a work repository, and check Settings again after major updates, because the run mode menu has changed more than once.
One more fact shapes the picture. The Auto-review classifier runs on Cursor's backend using small Cursor-managed models, so the actions it reviews leave your Mac.
04Can Cursor steal your code?
Two documented routes exist, and both start with an attacker.
A tampered extension on Cursor's marketplace tried to install a persistent SSH backdoor. On macOS, persistence used a LaunchAgent, a background job that starts at login, and it retried about every 60 seconds. Cursor's advice: uninstalling the extension may not be sufficient, and you should rotate credentials from a trusted system.
The second route is prompt injection, hostile text the agent reads and treats as instructions. In Cursor versions before 1.3.9, an attacker could chain indirect prompt injection to write .cursor/mcp.json and trigger remote code execution. The cause was that the agent could create new files inside the workspace, dotfiles included, with no approval. Updating past that version closes the hole.
05What has gone wrong in practice
Three examples. The full advisory list is on Cursor security.
- In a forum thread, a user running YOLO said Cursor deleted everything on their computer. Cursor staff replied by asking whether file deletion protection was enabled in the Auto-run settings, in June 2025.
- In July 2025, Backslash Security found four ways to bypass the Cursor denylist and said it cannot be relied upon. That was the old auto-run design, and current run modes differ.
- Cursor's GitHub advisories include CVE-2026-50549, rated critical: a sandbox escape via a symlink and failed path canonicalization.
Cursor's sandbox limits what a command can touch, using Seatbelt on macOS. Escapes like the one above mean it reduces risk and does not remove it. That is why the page asks you to stack settings instead of trusting one. The YOLO story has its own page: Cursor YOLO mode.
06Five settings that make Cursor safer
So is Cursor safe on your Mac? It is when these five are set, and each is either a setting or a hooks entry.
- Use Auto-review or Allowlist, not Run Everything. Run Everything runs every tool call automatically. Auto-review is the mode Cursor recommends.
- Turn on the protections. File-Deletion Protection stops the agent from automatically deleting files, including with
rm. External-File Protection stops automatic changes outside the workspace. Browser Protection stops the agent from automatically running browser tools. - Use the sandbox and keep the network closed. On macOS it uses Seatbelt through
sandbox-exec, with no setup, on Cursor v2.0 or later. Network access is blocked unless your network mode and allowlist open it, and Allow All lets the sandbox reach everything. Leave Allow All off. Note that the default allowlist includes github.com. Details: Cursor sandbox. - Narrow what it can read. Switch read access from System to Workspace if you run Cursor 3.23 or later. For untrusted repositories, use a basic text editor instead.
- Keep MCP approvals on. Do not click through a server you do not know. Your own hooks can add a check:
beforeShellExecutioncontrols shell commands, and exit code 2 blocks the action. Other exit codes fail open, so setfailClosed: trueon security-critical hooks.
Because agents can edit workspace files without approval, commit before every long session so you can undo what it did. Local sandbox config cannot weaken team-admin policies or hardcoded rules. Team admins can also enforce a command denylist in the Cursor CLI.
07So, is Cursor safe?
Is Cursor safe for everyday work? Yes, when you choose the mode and protections yourself. The vendor side is documented: certifications, Privacy Mode and public advisories. The open risks are the ones every agent shares: commands that run with your rights, extensions you install and text the agent reads from outside. Pick Auto-review, turn on the protections, keep the network closed and commit before long sessions.
Cursor's settings decide what the agent may run; they don't show what actually ran on your Mac. Artificial Immunity's free, open-source core watches Cursor, VS Code tasks and every program start, and reports what it finds without blocking anything.
Read the setup guide on how to run agents safely and the wider view in AI agent security. Or go back to the home page.
FAQ
Is Cursor safe?
For everyday work, yes, with the right settings. Cursor recommends the Auto-review run mode and has certifications and an audit attestation behind it. Its docs still describe run modes as best-effort guardrails, not a hard security boundary.
Is my data safe with Cursor?
With Privacy Mode on, Cursor will not train on your data, and it is available to anyone, free or Pro. Auto-review sends checks to a classifier that runs on Cursor's backend with small Cursor-managed models.
Can Cursor steal your code?
Two documented routes exist, and both start with an attacker. A tampered marketplace extension tried to install a persistent SSH backdoor. In versions before 1.3.9, a chain of prompt injection could write .cursor/mcp.json and trigger remote code execution, meaning the attacker runs commands on your machine.
Is Cursor at risk?
Like any editor with an agent, it has had security advisories. Examples are a critical sandbox escape via symlink, published 2026-06-05, and a sandbox escape through git hooks, fixed in 2.5.
Is Cursor legitimate?
Yes in the sense that matters for trust: it publishes advisories on GitHub and holds independent security certifications.
Can Cursor delete my files?
It can if you let it run commands. One forum user running the old run-without-asking setup, nicknamed YOLO, said Cursor deleted everything on their computer. File-Deletion Protection prevents the agent from automatically deleting files, including with rm.
Related guides
Sources
- Cursor docs, run modes
- Cursor docs, agent security
- Cursor docs, hooks
- Cursor security page
- Cursor blog, extension security notice
- Cursor GitHub security advisories
- NVD, CVE-2025-54135
- NVD, CVE-2025-54136
- Cursor forum, thread 103131
- Cursor forum, thread 98069
- The Register, Cursor safeguards bypassed
- Artificial Immunity