Vercel is challenging security researchers to crack its AI agent sandbox. Through HackerOne, the company is offering up to $1 million over two weeks. Each vulnerability can pay up to $50,000. The challenge runs from August 18 through September 1, 2026.
The idea behind the challenge is for researchers to simulate what malicious code would do. Anyone who succeeds in breaking out of the sandbox, gaining access to another user’s sandbox, or circumventing network restrictions can expect a payout.
AI agents read repositories, execute commands, install packages, and generate programs independently. Vercel Sandbox must run that code in an isolated environment and restrict access to the host, other sandboxes, and unauthorized network destinations, even if the executed code itself turns out to be malicious.
In addition to the compute side, network security is tested. Researchers may attempt to bypass the firewall, reach prohibited destinations, exfiltrate data, or obtain access credentials.
Rewards of up to $50,000
The reward amount depends on the severity and demonstrable impact. Critical vulnerabilities yield $25,000 to $50,000, while high-severity vulnerabilities yield $10,000 to $25,000. Moderate severity yields $5,000 to $10,000, and low severity yields $1,000 to $5,000. Reports must be submitted via HackerOne and include a working proof of concept. Findings based solely on static analysis do not count.
Vercel will review reports during the challenge and in the week following it, implement fixes, and pay out rewards once findings are confirmed. The company will use the discovered techniques to further refine Sandbox. Afterward, a report will be published detailing the attack techniques identified and the improvements implemented.