AI Security

Is Cursor Safe? Privacy Mode, Model Providers, Cloud Agents and the Security Advisories

Cursor sends your code to cloud models, so whether it is safe depends on one setting, the plan behind your account and how much freedom you give its agents. This guide sets out what Cursor documents today: what Privacy Mode guarantees, when code can be used for training, which providers see it, what is stored for search and for Cloud Agents, the MCP and auto-run risks, the security advisories with dates, and the controls on Teams and Enterprise.

Is Cursor safe? With Privacy Mode on, Cursor says your code is not used for training and its model providers store nothing; with it off, Cursor may store your code and train on it. Either way the code leaves your machine, because every request passes through Cursor's servers. The larger risk today is agents: they edit files, run commands and call MCP tools, and Cursor has published 33 security advisories.

Cursor is a code editor built around cloud models, so “is it safe” is two questions. One is about data: where your source code goes, who can keep it and whether it trains anything. The other is about behaviour: what the agent can do on your machine or in a cloud sandbox when it reads something hostile.

This review reflects Cursor’s website, documentation and GitHub advisories as we found them on 14 November 2026. Plan names changed this year too. The pricing page now lists Free, Plus, Premium, Super and Ultra for individuals, with Teams and Enterprise for organisations, and sells them alongside X and Grok features; customers on older plans such as Hobby and Pro keep them. What each plan costs is covered in our guide to Cursor’s plans and usage charges.

Is Cursor safe to use? Plans and data handling compared

PlanCan code be used for training?Opt-outWhat is storedHuman review
Free, Plus, Premium, Super, UltraYes, when Privacy Mode is offPrivacy Mode, under General in settingsTemporary encrypted file caches; Cloud Agent history and snapshots if you use themProviders may store and investigate prompts that trigger abuse detectors
TeamsNo by default: Privacy Mode is enabled for all team membersAdmins can enforce the settingAs above, plus custom Chat data retentionAs above
EnterpriseNo by default: Privacy Mode is onAdmins can enforce the settingAs above; Cloud Agent data can be capped at 90 daysAs above
PlanWhere it is processedAdmin controls and sign-inContract and assurance
Individual plansServers in various jurisdictions, including the United StatesNoneTerms of service and the privacy policy of Anysphere, Inc.
TeamsThe sameSAML or OIDC single sign-on, team-wide privacy settings, usage analyticsData processing agreement
EnterpriseUS-only data residency by enrolment; EU support is in developmentAdds SCIM, audit logs, and controls over auto-run, network, repositories, models and MCPData processing agreement; customer-managed keys for Cloud Agent data

Cursor’s security page, last updated on 25 August 2026, says the company holds AIUC-1, ISO/IEC 27001:2022 and ISO/IEC 42001:2023 certifications and a SOC 2 Type II attestation, with reports available through its trust portal, and commits to third-party penetration testing at least once a year. It also says Cursor uses no infrastructure in China and no subprocessors headquartered there. The pages we read make no statement about HIPAA or business associate agreements.

Cursor Privacy Mode: what it guarantees

Privacy Mode is one switch with large consequences. Cursor’s data use page, last updated on 3 September 2026, describes both positions.

With Privacy Mode on, the page says:

  • Customer data will not be used for training by Cursor.
  • Cursor maintains zero data retention agreements with all providers, and AI model providers will not store or train on your data.
  • Models that fall outside those agreements are labelled as such or need an admin to opt in for the workspace.

The exceptions are written into the same paragraph. Model providers, Cursor included, may run risk classifiers to detect violations of their usage policies. If a prompt or conversation triggers an abuse detector, the data may be stored for investigation and deleted under the provider’s own retention policy. So “zero retention” has a safety carve-out, as it does at most vendors.

Two notes apply in either position:

  • Requests go through Cursor’s backend even if you use your own API key, because that is where the final prompt is built.
  • Cursor temporarily caches file contents on its servers to cut latency. The files are encrypted with keys generated on your machine, the keys exist on Cursor’s servers only for the length of a request, and the cached content is not used for training when Privacy Mode is on.

To turn it on, open Cursor Settings, choose General and enable Privacy Mode. The security page says it is available to anyone, on free and paid plans, and that new team members inherit the team’s setting. An older variant, Privacy Mode (Legacy), blocks cloud storage altogether and for that reason cannot be used with Cloud Agents.

Does Cursor train on your code?

If Privacy Mode is off, it can. The data use page says that in this case Cursor may store and use codebase data, prompts, editor actions and code snippets to improve its AI features and to train its models. It adds that some inference providers may temporarily access and store model inputs and outputs to improve inference performance, and that this data is deleted after use.

Which position you start in depends on the account:

  • Teams: Cursor’s help page says Privacy Mode is enabled by default for all team members.
  • Enterprise: on by default.
  • Individual plans: the pages we read do not state a default. Check the setting on every machine and account before opening a work repository.

For an employer, the weak point is the personal account. A developer who signs in to Cursor with a private login on a company laptop decides the setting alone. Cursor’s answer is an MDM policy called Allowed Team IDs, which blocks personal accounts on corporate devices so that the team’s enforced Privacy Mode always applies.

Which model providers see your code?

Cursor is a router as well as an editor. Its enterprise documentation says prompts and code context are sent to language model providers, naming OpenAI, Anthropic and Google as examples, and that requests to Cursor’s own models may be processed by its inference providers. The data use page links to retention documentation from SpaceXAI, OpenAI, Anthropic and Meta. The full subprocessor list sits on the trust portal.

Three things follow for a security review:

  1. The provider depends on the model picked. A developer who changes model changes who processes the code. Enterprise admins can restrict the list with model access controls.
  2. A few models require retention. The documentation names Claude Fable 5.2, 5.1 and 5: Anthropic stores their inputs and outputs for automatic and human harm-prevention review, and says it does not use them for training. For Enterprise customers and anyone with Privacy Mode on, requests to these models fail until an admin approves the retention policy in the dashboard, and the approval applies to the whole team.
  3. Your own API key changes the contract. With a personal key, zero data retention depends on your agreement with that provider, not on Cursor’s. Cursor’s hardening guide tells organisations that rely on its agreements to restrict personal keys.

Location is narrower than the provider list suggests. Enterprise customers can enrol a team in US-only data residency, which keeps inference, processing and storage in the United States for supported models and features at a 10% uplift on model pricing. Cursor says EU support is in development, so an organisation that needs processing inside the EU cannot get it from Cursor today.

Codebase indexing: what leaves your machine

Search is where Cursor’s documentation has changed most. Older descriptions, including ones still widely quoted, say Cursor uploads chunks of your code to compute embeddings and stores those embeddings with obfuscated file paths. The current documentation for Agent search says something different. Its privacy note on the built-in search engine, Instant Grep, says the index is built and queried on your own machine, that Cursor uploads neither file paths nor code to build a search index, and that it stores no embeddings of your codebase for search.

That does not make search private in the sense that matters most: when the agent opens a matching file, that content can be included in the model request, under the terms described above. If you run an older client, check which behaviour your version has.

To keep specific files out of the agent’s reach, Cursor reads a .cursorignore file in the project root, written in .gitignore syntax:

.env*
secrets/
*.pem
customer-data/

Read the limits before relying on it. The documentation says terminal commands and MCP server tools used by the agent cannot be blocked from code governed by .cursorignore, and that complete protection cannot be guaranteed because language models are unpredictable. It is a guardrail. Secrets should not be in the working tree at all.

Cloud Agents: the feature that stores your code

Cursor’s enterprise documentation is blunt: “Cloud Agents are the only feature that requires Cursor to store code.” They are available from the Premium plan upward and work with Privacy Mode on, in which case Cursor says it never trains on the code and retains it only to run the agent. What a review needs to know:

  • Where they run. Your code runs inside Cursor’s AWS infrastructure in isolated virtual machines and is stored on the VM disks while the agent is accessible. Cursor’s GitHub app needs read-write access to the repositories you want changed.
  • What is kept. Conversation history is kept indefinitely by default. Environment snapshots, which are encrypted copies of the VM disk, are deleted after 30 days without use, and each start or resume resets the 30 days. Snapshots cannot be deleted on demand. Enterprise admins can cap conversation data at 90 days.
  • Commands run unattended. The documentation says the agent auto-runs all terminal commands, and states the consequence itself: auto-running creates a risk of data exfiltration. A prompt injection could trick the agent into uploading code to a hostile site.
  • Internet access is on by default. Egress can be limited to an allowlist per user, environment or team, and Enterprise admins can lock the team setting.
  • Secrets have three types. Environment variables are visible to the agent. Runtime Secrets are redacted from tool results, transcripts and commits but still readable from the agent’s terminal. Build Secrets never reach the running agent.

Cloud Agents are optional. Cursor’s own advice is that if your security policy prohibits code storage, you should not enable them. Enterprise customers can encrypt Cloud Agent data with their own keys.

MCP, auto-run and prompt injection risks

On your own machine, the agent’s defaults are conservative in one place and permissive in two. According to Cursor’s agent security page:

  • Reading files and searching code needs no approval.
  • Editing workspace files needs no approval, except configuration files. Changes are written to disk at once, so version control is your undo. Cursor warns that with auto-reload enabled, an agent’s changes might execute before you review them.
  • Terminal commands need approval by default. Run Modes can let trusted calls through, from a simple allowlist to an Auto-review classifier. Cursor describes these as best-effort guardrails, not a hard security boundary.
  • MCP connections need approval, and so does each tool call, unless you pre-approve tools with an allowlist.
  • Network requests are limited to GitHub, direct link retrieval and web search providers under default settings.
  • Workspace trust is disabled by default. Enable it in settings.json if you open repositories you did not write:
{
  "security.workspace.trust.enabled": true
}

MCP servers deserve their own review. Cursor’s guidance is to install them only from trusted developers, check what data and APIs they reach, use restricted API keys and read the source of critical ones, because a server can reach external services and run code on your behalf. A repository, a web page or an MCP tool result can all carry instructions written for the agent. Our prompt injection prevention guide explains the pattern, and our review of MCP servers for enterprise use covers how to vet one.

Cursor security advisories: what has been fixed

Cursor publishes advisories in its GitHub repository. We counted 33 on 14 November 2026. The most recent show where the weak points have been:

PublishedAdvisorySeverityDetail
14 July 2026Sandbox escape via tampered Python virtual environments; sandbox escape via privileged containersHighTwo advisories on the same day; the first is fixed in 3.1.2, the second affects versions before 3.0
6 July 2026Cloud Agent browser sandbox escape, CVE-2026-61613HighWeb content loaded in a browser-enabled Cloud Agent could reach a local agent endpoint and expose files, environment variables and GitHub App tokens. Fixed on Cursor’s side on 31 March 2026
5 June 2026Desktop sandbox escape via symlink, CVE-2026-50549; sandbox escape via agent-controlled working directoryCriticalA malicious agent could write outside the workspace and reach unsandboxed code execution with no interaction beyond a benign prompt. Fixed in version 3.0
21 May 2026Desktop sandbox escape via Claude hook configurationHigh
9 March 2026Arbitrary code execution via prompt injection and allowlist bypassHighA malicious web page could get commands run even with Auto-Run set to the allowlist
13 February 2026Sandbox escape via Git hooksHigh
14 January 2026Terminal tool allowlist bypass via environment variablesHigh
3 November 2025Speedbump modal bypass in the MCP server deep-link; two sensitive-file protection bypasses on WindowsHighThree advisories on the same day

Cursor fixes and discloses, which is what you want from a vendor. The pattern is also consistent: most advisories are ways for a misled agent to get out of its sandbox or past an approval prompt. Keep the client updated, treat versions before 3.1.2 as unsafe, and assume the approval prompts can fail.

Is Cursor safe for work? Team and Enterprise controls

Cursor’s hardening guide describes a shared responsibility: Cursor secures the platform, and the customer configures identity, privacy enforcement, agent controls and monitoring. Most of the levers that enforce anything across an organisation are Enterprise features.

  • Teams gives central billing and administration, usage analytics, team-wide privacy settings, custom Chat data retention and SAML or OIDC single sign-on.
  • Enterprise adds SCIM seat management, audit logs and service accounts, group-level settings, auto-run, browser and network controls, and repository, model and MCP access controls. The MCP control is an allowlist of servers by command or URL pattern, with optional per-tool lists.

Cursor’s own “do these first” list for administrators is a good summary of what it considers unsafe by default: enforce Privacy Mode for the organisation, set the Run Mode policy to Auto-review and not Run Everything, enable sandboxing, apply network allowlisting, review plugins and MCP servers, add .cursorignore entries for secrets, lock identity with single sign-on, SCIM and Allowed Team IDs, restrict personal API keys, and stream audit logs to a SIEM.

What developers and IT should do

If you write code in Cursor:

  1. Confirm Privacy Mode is on before opening any repository you do not own personally.
  2. Use your work account, never a personal one, for work code.
  3. Keep terminal approval on. Do not pick a run-everything mode on a machine that holds credentials.
  4. Open unfamiliar repositories with workspace trust enabled, and read a project’s rules and MCP configuration before letting the agent act on them.
  5. Keep secrets out of the tree, add .cursorignore as a second layer, and update the client when a release is offered.

If you run engineering or security:

  1. Buy the plan that can enforce your policy. If you need audit logs, SCIM or MCP allowlists, that is Enterprise.
  2. Enforce Privacy Mode and deploy Allowed Team IDs through MDM.
  3. Decide which models are allowed, whether any retention-required model is approved, and whether personal API keys are permitted.
  4. Decide on Cloud Agents explicitly. If enabled, restrict egress to an allowlist, scope the GitHub app to named repositories and set the retention cap.
  5. Write the result into your AI acceptable use policy, including which repositories may never be opened in a cloud assistant.

Other coding assistants raise the same questions under different terms. Our Claude guide includes how Claude Code handles code and telemetry, and the GitHub Copilot alternatives page covers the move to a self-hosted assistant.

When code must not leave your network

Cursor has no offline or on-premises mode: the data use page is explicit that requests go through its backend, even with your own API key. Privacy Mode, enforced, is a strong contractual position for ordinary commercial code. It does not answer obligations that forbid third-party processing, such as some defence, banking and client-escrow terms, and it cannot place inference in the EU today.

For that code the choice is an assistant whose model runs on hardware you control. Our comparison of private and public LLMs covers the general trade-off, the local coding model guide covers which open-weight models are worth testing, and our on-premises code assistant round-up covers the tools.

How VDF AI fits

VDF AI does not run, host, resell or integrate with Cursor. Its own product for developers is VDF Code, a coding assistant for enterprise teams that provides completions, chat and pull request review and is deployed in the organisation’s VPC, in its data centre or fully air-gapped. It works through extensions for VS Code, JetBrains IDEs, Neovim and Visual Studio, retrieves context only from repositories the organisation authorises, and uses the models the organisation approves.

The difference from a single developer’s Cursor set-up is where the decisions sit. In an on-premises VDF Code deployment, prompts, embeddings and telemetry stay inside the network, and the model, the repositories in scope and the access rules are set centrally instead of per laptop. To see it against your own repositories, book a demo.

Sources

Cursor’s own pages and advisories, consulted on 14 November 2026.


Need a coding assistant that keeps source code inside your network? See how VDF Code runs on models you host, or book a demo.

Frequently asked questions

Is Cursor safe to use?

Cursor is a mainstream editor with published security practices, and it is reasonable to use with care. It is not a local tool: every AI request goes through Cursor's backend to a cloud model, even with your own API key. With Privacy Mode on, Cursor says your code is not used for training and providers do not store it. With it off, Cursor may store and train on code data. Agents add a second risk, because they can edit files and run commands.

Does Cursor train on your code?

Only when Privacy Mode is off. Cursor's data use page says that in that case it may use and store codebase data, prompts, editor actions and code snippets to improve its AI features and train its models. With Privacy Mode on, it says customer data will not be used for training by Cursor and that model providers will not store or train on it. Privacy Mode is on by default for Teams and Enterprise; check the setting yourself on an individual plan.

What does Cursor Privacy Mode do?

Privacy Mode is a setting under General in Cursor's settings. When it is on, Cursor says it will not train on your data and that it holds zero data retention agreements with its model providers. It does not stop code being sent to the cloud, and it has documented exceptions: providers may store prompts that trigger abuse detectors, a few models require retention and need admin approval, and Cloud Agents store repository copies while they run.

Is Cursor safe for work?

On a Teams or Enterprise plan with Privacy Mode enforced, it is suitable for most commercial code. Teams adds SAML or OIDC single sign-on and team-wide privacy settings. Enterprise adds SCIM, audit logs, controls over auto-run, network, repositories, models and MCP servers, and optional US-only data residency. A personal account used for work gives the employer no control over the setting, which is why Cursor offers a device policy to block personal logins.

Is Cursor safe for proprietary code?

It depends on what your obligations say. With Privacy Mode enforced, the vendor's commitments are no training and no retention by model providers, under a data processing agreement. The code still leaves your network and is processed by Cursor and its providers. If a contract, export rule or regulator says the code may not be processed by third parties, no Cursor setting satisfies that, and a self-hosted assistant is the option.

Can prompt injection make Cursor run commands?

It has in the past, and Cursor's own documentation treats it as a live risk. The foreground agent asks before running terminal commands by default, but Run Modes can relax that, and Cursor calls those modes best-effort guardrails, not a security boundary. Cloud Agents run every command without asking. Cursor's GitHub repository listed 33 security advisories when we counted, including two critical sandbox escapes published on 5 June 2026 and fixed in version 3.0.

Filed under
CursorAI coding assistantsAI securityAI governancedata sovereigntyprivate AI
AI Governance

Is your AI governance audit-ready?

Get a readiness review of your AI controls — policy, oversight, audit trails, and EU AI Act evidence — mapped against what production actually requires.

Keep reading