PwnEscape lets Linux guests expose host memory through a kernel networking flaw
Researchers at pwn.ai say CVE-2026-99999 lets guest-provided checksum metadata overwrite an IPv4 header that the Linux host has already checked, leading to a repeatable 32-byte host-memory disclosure. Pwn.ai demonstrated host code execution on its own KVM setup, while SevyDevy reports a production memory leak in Vercel sandboxes but no production host takeover. Linux merged a fix that checks checksum positions before completion in five IPv4 processing paths.
Bottom line — The upstream fix is merged; Linux hosts running untrusted guests need a patched kernel to close the 32-byte disclosure path.
Go deeper 5
-
Pwn.ai traced the flaw to deferred checksum processing in the Linux networking stack, not to a Vercel-specific configuration.
-
The 32-byte disclosure was demonstrated in controlled tests; pwn.ai says its separate host-execution chain required additional memory manipulation.
-
SevyDevy says Vercel reported no evidence of customer-data access in its sandbox.
-
Linux merged the guard on 14 October 2026, according to pwn.ai; the patch is described as defence in depth, not a complete redesign of checksum validation.
-
The Linux CNA rated the flaw 9.3 on CVSS, while Red Hat assigned 8.8, according to DEV Community.