A critical vulnerability in the Linux kernel has been recently disclosed, allowing code within a virtual machine to access memory on the host computer. This flaw, known as CVE-2026-99999 or PwnEscape, resides in the Linux kernel’s networking code, posing a risk to shared infrastructure across cloud providers.
Discovered during a security challenge by Vercel, the flaw was demonstrated against Vercel’s production sandboxes. The exploit, showcased by Paulos Yibelo of pwn.ai, involves sending specially crafted network packets from a virtual machine to read the host’s memory, crash the host, and even execute code on it.
Vercel confirmed the memory leak on its sandboxes, where each packet sent by the guest could retrieve approximately 32 bytes of the host’s memory. While the crash and code execution were observed in a controlled environment, reaching administrator control of the host highlighted the severity of the vulnerability.
The flaw has been rated 9.3 out of 10 in severity and designated as local, requiring the attacker to have existing code running within a guest. Although the fix has been implemented in the Linux kernel, it is essential for users to update their kernels to the patched versions to mitigate the risk.
As the fix is a partial measure, additional security measures are recommended, such as limiting the potential impact of a breach and ensuring host kernels are up to date. Vercel’s proactive approach in keeping host kernels updated and implementing security measures like Firecracker’s jailer showcases best practices in mitigating such vulnerabilities.
The discovery of this flaw in Vercel’s $1 million challenge underscores the importance of rigorous security testing and the role of AI-assisted research systems in identifying and addressing critical vulnerabilities. Researchers leveraging AI models to analyze source code and identify vulnerabilities demonstrate the evolving landscape of cybersecurity.
