Sandbox Escape Vulnerability in Skyvern AI Product

CVE-2026-82447

What is CVE-2026-82447?

Skyvern AI prior to version 1.0.45 is affected by a sandbox escape vulnerability specifically within the TextPromptBlock component. The flaw allows an attacker to manipulate workflow parameters or upstream block output to inject malicious Jinja template syntax. This results in the execution of arbitrary code with server process privileges. The vulnerability stems from the dual rendering of prompts—initially through a sandboxed Jinja environment followed by an unsandboxed environment—thereby bypassing security controls designed to restrict untrusted code execution.

References