PwnEscape, CVE-2026-99999 (CVSS 9.3): the Vercel Sandbox 'KVM escape' is a Linux kernel IPv4 checksum bug — a VM guest can read host memory (shown live on Vercel), crash the host, and in the researcher's lab run code as root on it; fixed in stable kernels 5.10.272 to 7.2.11
Update 16 Oct: the guest-to-host escape Vercel's CEO confirmed on 3 October now has a CVE and a public write-up. Researcher Paulos Yibelo of pwn.ai published 'PwnEscape' on 15 October and the Linux kernel CNA assigned CVE-2026-99999, rated 9.3 Critical (CVSS 3.1, local attack vector, scope changed); NVD had received but not yet analysed it. The bug is not in KVM or Firecracker but in the host kernel's IPv4 networking code: checksum metadata supplied by a guest could point into the IPv4 header, so skb_checksum_help could rewrite the header after the kernel had already validated it.
Vercel's own findings post (14 Oct) says the host-memory leak was demonstrated against its production Sandbox: each packet a guest sent could carry back about 32 bytes of host memory, repeatably. Crashing the host with two crafted packets and getting root on the host were shown only on the researcher's own machines built to match Vercel's; he stresses this is not a claim of root in any provider's production.
Vercel says no report showed anyone reading customer files, secrets or app data, and called it the challenge's only Critical finding. Because most clouds isolate tenants on the same Linux host networking layer, the exposure is broad for anyone running untrusted guests. The fix is a single commit (6785011f8b16) by Red Hat's Michael S. Tsirkin, which its author calls a defence-in-depth check; fuller input validation is still being worked on.
An attacker must already run code inside a guest; it is not reachable from the open internet, and there is no known in-the-wild use or CISA KEV listing. pwn.ai says the research used its Omega-Pwn harness with an AI model as the main reasoning engine.
Earlier wire-only reporting (Rauch X post 3 Oct, Tech Insider and CSN 4 Oct) said Vercel paid US$50,000.
UPDATE (20 Oct): Red Hat rates CVE-2026-99999 Important, CVSS 3.1 8.8 (it scores privileges required as low, unlike the CNA's 9.3), lists every tracked RHEL stream from 6 to 10 (kernel and kernel-rt), RHEL for NVIDIA and OpenShift 4 CoreOS as affected with fixes still in progress, and says no acceptable mitigation exists.
WA's Cyber Security Unit issued advisory 20261019001 on 19 October and has seen no exploitation on WA Government networks.
Primary: pwn.ai write-up and kernel commit; vendor: Vercel; wire: The Hacker News (16 Oct).
- Product
- Linux kernel IPv4 networking (host side) — affects KVM/Firecracker and other hypervisor hosts running untrusted guests; found via Vercel Sandbox
- Versions
- Fixed (first fixed per CVE record): 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.57, 7.2.11 and mainline commit 6785011f8b16. Distribution kernels may backport without changing the version number — check your vendor advisory.
- CVSS
- (CVSS 3.1, Linux kernel CNA; NVD analysis pending)
CVSS:3.1/AV:L/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H - Exploited in Australia?
- unknown
- Patch to
- Update the HOST kernel to a fixed stable release or your distribution's patched build and reboot into it; patching guests alone does not help. Red Hat had no fixed RHEL or OpenShift kernel as of 20 Oct; watch access.redhat.com/security/cve/cve-2026-99999 and patch hypervisor hosts first when errata land. Prioritise multi-tenant hypervisors, CI runners and AI-agent sandboxes that run untrusted code. No separate workaround was published.
- Primary: pwn.ai — PwnEscape: guest-to-host KVM escape (CVE-2026-99999) (15 Oct 2026); fix: Linux commit 6785011f8b16
- Vendor: Vercel — Findings from the $1 million Vercel Sandbox challenge (14 Oct 2026); CVE record: nvd.nist.gov/vuln/detail/CVE-2026-99999
- CVE: CVE-2026-99999
- Source: The Hacker News — Critical PwnEscape Linux kernel vulnerability lets guest VMs read host memory (16 Oct 2026)
