Parallels Desktop Flaw Lets Unprivileged Mac Users Gain Root Access
JFrog's 'ParaShells' research details CVE-2026-90894, a Parallels Desktop local privilege escalation letting unprivileged macOS users gain root via tar argument injection.
JFrog researcher Yuval Moravchick disclosed CVE-2026-90894 ('ParaShells'), a critical local privilege escalation flaw demonstrated against Parallels Desktop 26.4.0 build 57513 on Apple silicon. The world-writable /var/run/prl_disp_service.socket accepts connections from unsigned clients, and argument injection into a tar extraction command via the --use-compress-program option executes attacker-controlled code with root privileges. The attack needs no administrator permissions, running VM, or Parallels-signed client, and a one-line PoC yielded a root shell. Risk is highest on developer workstations, shared training systems, and enterprise Macs.
- CVE-2026-90894 lets unprivileged local users execute code as root
- World-writable prl_disp_service.socket accepts unsigned client connections
- tar --use-compress-program argument injection executes attacker code with root privileges
- No admin rights, running VM, or signed Parallels app required
- Exploitable via compromised npm packages, malicious installers, or poisoned CI workflows
- CVEs
- CVE-2026-90894
- Products
- Parallels Desktop
Vulnerabilities mentioned
All| CVE | Vulnerability | CVSS | EPSS | Signals | Published |
|---|---|---|---|---|---|
| CVE-2026-90894 | Local Root Privilege Escalation in Parallels Desktop via tar Argument Injection | 7.8 | — |
From the original article
2 min read at gbhackers.com
A critical local privilege escalation vulnerability in Parallels Desktop could allow an unprivileged macOS user or a malicious process to gain root-level access to the host system. The issue, tracked as CVE-2026-90894, was disclosed by Yuval Moravchick from JFrog’s Vulnerability Research team and has been named “ParaShells.” This flaw was demonstrated against Parallels Desktop version […]
By Divya · gbhackers.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.