Sandbox vulnerability gives rogue AI agents a path to escape and leak data, research says

- Researchers at AIR disclosed a certificate-handling vulnerability in Vercel Sandbox that lets rogue AI agents bypass network allowlists by crafting certificates that cause an approved server to relay requests to attacker-controlled servers, enabling data exfiltration without the sandbox directly contacting unauthorized destinations. AIR notified Vercel, no public evidence of exploitation has been found, and the issue is described as difficult to patch and potentially longstanding.