Cursor
Security FAQ

What vulnerabilities are found in Cursor apps?

Get instant answers about your app's security.

Short Answer

Scans of Cursor apps surface a recurring set of findings: prompt injection in mcp servers, workspace trust exploitation, and code suggestion security flaws. 4 of these have documented real-world exploitation.

Detailed Answer

The vulnerabilities actually found in Cursor apps

Not theoretical OWASP categories — specifically what appears when VAS, security researchers, and bug bounty hunters look at live Cursor deployments:

  1. **[CRITICAL]** **Prompt Injection in MCP Servers** *(medium likelihood)*

Malicious content in MCP tool responses can execute arbitrary commands.

*Observed:* CVE-2025-54135 (CurXecute): Slack messages could trigger remote code execution.

*Fix:* Review MCP server sources. Avoid untrusted MCP integrations. Watch for suspicious tool calls.

2. **[CRITICAL]** **Workspace Trust Exploitation** *(medium likelihood)*

Malicious .cursor/rules files execute when opening untrusted projects.

*Observed:* The 'Rules File Backdoor' vulnerability affects both Cursor and Copilot.

*Fix:* Enable Workspace Trust in settings. Review .cursor/ files before opening projects.

3. **[HIGH]** **Code Suggestion Security Flaws** *(high likelihood)*

AI suggests vulnerable patterns: SQL injection, hardcoded secrets, weak auth.

*Observed:* Research shows ~40% of AI code suggestions contain security issues.

*Fix:* Review all AI suggestions critically. Run security scans on generated code.

4. **[MEDIUM]** **Privacy and Code Exfiltration** *(medium likelihood)*

Code sent to AI servers may expose proprietary logic or secrets.

*Fix:* Enable Privacy Mode. Use .cursorignore for sensitive files.

5. **[HIGH]** **Supply Chain via Package Hallucination** *(low likelihood)*

AI suggests non-existent packages that attackers could register.

*Observed:* LLMs hallucinate npm package names that attackers then create with malware.

*Fix:* Verify all package suggestions exist. Check package reputation before installing.

Distribution by severity

Of the findings above, 2 sit at critical impact (full data exposure), 2 at high (significant data or account compromise), and the rest are medium-or-lower (attack surface expansion). A first-scan Cursor app typically has 2–4 findings from this list live at any moment.

How to know which ones are in your app

Run a VAS scan. Each finding above is tested directly — we query your database to verify access controls are active, scan bundles for key patterns, probe auth endpoints for rate limiting, and check security headers in live responses. Output is a per-finding report with evidence and fix.

Security Research & Statistics

10.3%

of Lovable applications (170 out of 1,645) had exposed user data in the CVE-2025-48757 incident

Source: CVE-2025-48757 security advisory

4.45 million USD

average cost of a data breach in 2023

Source: IBM Cost of a Data Breach Report 2023

500,000+

developers using vibe coding platforms like Lovable, Bolt, and Replit

Source: Combined platform statistics 2024-2025

Expert Perspectives

There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.

Andrej KarpathyFormer Tesla AI Director, OpenAI Co-founder

It's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

Andrej KarpathyFormer Tesla AI Director, OpenAI Co-founder

Check Your Cursor App's Security

VAS scans for all the security issues mentioned above. Get a comprehensive security report in minutes.

Get Starter Scan

More Questions About This Topic

How severe are the vulnerabilities typically found in Cursor apps?

Cursor apps lean critical: Prompt Injection in MCP Servers alone can expose the full user dataset in one query. That's a critical-impact finding with medium likelihood — it is, in practice, the default state of an unscanned Cursor app. Compare to e.g. missing security headers (medium) which require additional exploitation. Triage accordingly: critical findings are measured in minutes-to-breach, mediums in weeks.

How do I fix vulnerabilities once they're found in my Cursor app?

Each finding comes with a specific fix. Example: for "Prompt Injection in MCP Servers" → Review MCP server sources. Avoid untrusted MCP integrations. Watch for suspicious tool calls. VAS exports these as markdown you can feed directly into Cursor's AI (or any other AI assistant) to apply the fix in-place. Re-scan afterward to confirm.

Can vulnerabilities in Cursor apps be exploited by a non-expert attacker?

Most can. Extracting an exposed API key is a single "view source" operation. Querying a table without RLS is a `curl` command. Exploiting missing rate limiting requires scripting skills equivalent to "follow a tutorial." Only a handful of the findings above (e.g., chained auth bypass) require specialist knowledge — the rest are routinely exploited by automated scanners with zero human involvement.