Vercel’s $1 Million Bug Bounty Challenge Uncovers Critical Linux Kernel Flaws
A recent bug bounty challenge run by cloud platform provider Vercel has uncovered two critical flaws in the Linux kernel’s networking stack, potentially impacting millions of users worldwide. The two-week challenge, which offered a $1 million reward, drew 1,285 reports from researchers and testers, but none were able to access customer data.
The bug bounty program was designed to test the security of Vercel’s sandbox environment, a Firecracker-based microVM that isolates untrusted AI-agent code. The challenge, which ran from August 18 to September 1, invited both black box (HackerOne) and white box (Trail of Bits engineers) testers to attempt to escape the sandbox.
The results are in, and while none of the submissions were able to gain access to customer data, several critical vulnerabilities were uncovered. The most significant finding was two independent defects in the Linux kernel’s networking stack, which could allow attackers to leak memory from the host’s kernel or crash the host deterministically. These flaws have far-reaching implications, as many major cloud providers use the same layer of the Linux kernel to isolate customer workloads.
The fact that these vulnerabilities were discovered during the bug bounty challenge is significant, as it means that Vercel was able to learn about them two weeks ahead of the kernel maintainers. This early warning will allow Vercel to address the issues before they can be exploited by attackers.
In addition to the critical flaws, the white box testing by Trail of Bits engineers produced 20 findings, but it was their architectural advice that proved most valuable. The engineers recommended a number of structural changes to improve the security of the sandbox environment, including “stop trusting the guest” – in other words, do not rely on the tenant input.
The sheer volume of reports received during the challenge presented a significant challenge for Vercel’s report triage team. Initially, each report was reviewed by a human, but as the volume increased, Vercel developed its own agentic triaging solution using an agent built on the Vercel Eve agent framework. This new agent reads each report, checks it against the program policy, and runs the researcher’s proof of concept inside a real Vercel Sandbox.
The success of this bug bounty challenge has implications beyond just Vercel’s platform. The volume of reports received is a testament to the power and speed of researchers working with AI-assistance, and highlights the need for companies to develop robust security measures to counter the increasing threat of AI-driven attacks.
As the security landscape continues to evolve at breakneck speed, it’s clear that human decision-making will inevitably add latency to the security response. In this instance, Vercel was able to remove the human from the loop and rely on its AI defense system to improve the speed of the process. This may be a lesson for the future – as everything speeds up, automated decision-making may become an essential component of effective security.
In conclusion, Vercel’s bug bounty challenge demonstrates the strength of its sandbox environment and highlights the importance of robust security measures in today’s fast-paced digital landscape. As companies continue to grapple with the challenges of AI-driven attacks, it’s clear that collaboration between humans and machines will be essential to staying ahead of the threats.
Source: SecurityWeek — 2026-09-15