Trust Issues: Bypassing Cursor AI's 'Safe Mode' via Shell Built-ins
Vulnerability ID: CVE-2026-22708
CVSS Score: 7.2
Published: 2026-01-14
A logic flaw in Cursor AI's terminal allowlist allowed attackers to execute arbitrary code by leveraging shell built-ins like 'export' and 'alias' to poison the environment, effectively bypassing security restrictions.
TL;DR
Cursor AI's 'Auto-Run' mode promised safety via a command allowlist. However, it only scrutinized external binaries, completely ignoring shell built-ins. By tricking the AI into running commands like export PATH=... or alias, an attacker could poison the shell environment. The next time a 'safe' command (like git) was executed, the shell would invoke a malicious payload instead, resulting in full RCE.
⚠️ Exploit Status: POC
Technical Details
- CWE ID: CWE-78 (OS Command Injection)
- Attack Vector: Network (Prompt Injection)
- CVSS Score: 7.2 (High)
- CVSS Vector: CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N/E:U
- EPSS Score: 0.00054 (Low Probability)
- Exploit Status: Proof of Concept (PoC)
- Affected Component: Cursor Agent / Terminal Tool
Affected Systems
- Cursor AI Code Editor (macOS)
- Cursor AI Code Editor (Linux)
- Cursor AI Code Editor (Windows)
-
Cursor: < 2.3 (Fixed in:
2.3)
Exploit Details
- Pillar Security: Research notes detailing the bypass of Cursor's allowlist via environment poisoning.
Mitigation Strategies
- Disable Auto-Run Mode in Cursor settings to enforce human-in-the-loop verification.
- Treat shell built-ins (export, alias, source) as high-risk commands.
- Isolate AI development environments using containers or VMs.
Remediation Steps:
- Update Cursor to version 2.3 or later via the internal updater or official website.
- Verify that the 'Allowlist' settings are correctly configured and not set to allow all.
- Review shell history (
~/.zsh_historyor~/.bash_history) for unexpected export or alias commands if you suspect compromised sessions.
References
Read the full report for CVE-2026-22708 on our website for more details including interactive diagrams and full exploit analysis.
Top comments (0)