Vercel Sandbox flaw lets rogue AI agents escape and leak data, researchers say

AI security startup AIR found a network filtering flaw in Vercel Sandbox that lets rogue agents bypass allowlists and exfiltrate data via a crafted certificate. Vercel acknowledged the issue but has no patch yet; no exploitation has been observed.

• Flaw in Vercel Sandbox lets agents bypass network restrictions using a crafted certificate
• Data reaches an attacker server via an approved server, so the sandbox sees no direct connection
• Vercel confirmed the issue but has no timeline for a fix
• AIR found no evidence of exploitation in the wild