A newly disclosed critical vulnerability in the Linux kernel, tracked as CVE-2026-99999 and nicknamed PwnEscape, allows code running inside a guest virtual machine to read memory from the host. The flaw sits in the kernel’s networking code and affects the shared layer used by cloud providers to isolate customer workloads.
The vulnerability was demonstrated against Vercel’s production sandbox environment, with Vercel confirming the findings privately and noting that the leak involved memory from the host being exposed back to the guest via specially crafted network packets. In the researcher’s tests, each packet could reveal around 32 bytes of host memory, and continued exploitation was possible as long as packets were sent.
The same flaw also enabled crashes and, in controlled conditions, attempted code execution on the host; however, the researchers emphasised these outcomes were lab results and not observed in production environments.
The issue has been assigned CVE-2026-99999 and rated 9.3/10 “Critical” by the kernel maintainers; the U.S. National Vulnerability Database had begun its review at the time of reporting. The recommended fix, implemented by Red Hat’s Michael S. Tsirkin and merged into mainline, adds checks to ensure packet checksum data cannot point into the IPv4 header, applied at five points in the networking code.
Stable branches backported the patch to versions 5.10, 5.15, 6.1, 6.6, 6.12, 6.18, 7.2, with corresponding first fixed versions listed. Organisations should update to the fixed kernels or vendor builds and reboot. Where immediate patching isn’t possible, Vercel’s approach—restricting nested virtualization and running microVMs under a jailer—helps limit potential impact, but does not remove the vulnerability.