Skip to content

Back to the catalog

lee1981 (@lee1981b) on X
x.comadvisory

lee1981 (@lee1981b) on X

CyberForge CVE of the Day #080 covers CVE-2026-99999 ("PwnEscape"), a Linux kernel IPv4 checksum-completion flaw. Crafted checksum metadata can trigger a deferred write into an already-validated IPv4 header, which allows host memory disclosure from a guest or microVM over virtio-net and TAP. It gives upstream fixed branches (5.10.272 to 7.2.11), Amazon Linux, Red Hat and Ubuntu status, CVSS 9.3, EPSS 0.211% and a SOC investigation trail. No IOCs were verified.

TL;DR Defender briefing on the PwnEscape Linux kernel IPv4 checksum bug that lets microVM guests reach host memory, with fixed versions and a hunt and verify trail.
Document

What it is

CyberForge CVE of the Day #080 on CVE-2026-99999 ("PwnEscape"), a Linux kernel IPv4 checksum-completion flaw. The Linux CNA titles it "ipv4: validate checksum_start before completing checksum" and scores it CVSS v3.1 9.3 (AV:L/AC:L/PR:N/S:C). The post is mainly a host-side scoping and investigation guide, not an exploit write-up.

How it works

Checksum offload defers work using metadata that describes where the checksum starts and where it goes. Less-trusted guest metadata can authorise a write into an IPv4 header that was already validated. Later fragmentation processing trusts the altered header. The upstream fix adds an IPv4 guard before completion at five consumers, with a shared helper in a sixth file. Upstream calls it defence in depth.

Preconditions / Affected

The defect is in the host kernel, not in KVM or the VMM binary. The post says to look at self-managed hosts running untrusted guests or microVMs over virtio-net and TAP, and to confirm the real backend, because a "KVM" label alone does not settle exploitability.

Upstream traces the introduction to Linux 2.6.27. Stable fixed boundaries:

  • 5.10.272
  • 5.15.223
  • 6.1.190
  • 6.6.159
  • 6.12.113
  • 6.18.57
  • 7.2.11
  • 7.3-rc7, the original-fix boundary (source history, not a production target)

Distribution backports can fix an older-looking base version, so match the exact vendor stream and package revision.

  • Amazon Linux: the tracker marks the rows Fixed. Four revisions differ between the ALAS advisories and the overview bulletin. For 5.15, ALAS2KERNEL-5.15-2026-117 lists kernel-5.15.221-157.262.amzn2, while the bulletin lists kernel-5.15.220-156.256.amzn2. Use the current advisory.
  • Red Hat: the captured API lists affected rows with no fixed package or erratum.
  • Ubuntu: Medium priority, with "Needs evaluation" for standard linux rows. A "Not in release" entry for one package does not clear the host.
  • AWS: describes the self-managed KVM/virtio-net scope as unpatched 4.4 to 6.18. EC2, Lambda and Fargate are stated as not affected, which does not clear a separately self-managed virtualisation layer.

The privilege assumption differs by source. The CNA models a workload with no host credentials sending through its virtual interface. AWS describes privileged access inside the guest.

What to do

  1. Prioritise self-managed hosts carrying untrusted guests, including standby nodes and boot images that can recreate the old state.
  2. Apply the supported distribution kernel update, not isolated cherry-picks, which the Linux CVE team does not recommend.
  3. Drain and reboot, then verify the running host kernel. An installed package does not show which kernel booted, and output from inside a guest describes only the guest.
  4. Update launch images and replacement capacity as well as live nodes.
  5. Before disruptive change, preserve kernel and boot records, VMM lifecycle, interface mappings and any retained crash evidence. Do not trigger a crash or replay packets.
  6. For the hunt, review existing journals and crash reports for faults in skb_checksum_help and ip_do_fragment. These are leads, not unique signatures. Join host, boot, time and guest placement.

Caveats

  • No workaround was verified. AWS says none exists, and Red Hat says mitigations are unavailable or fail its criteria. Admission restrictions and workload isolation reduce exposure but are not certified fixes.
  • Vercel confirmed host-memory disclosure on production sandboxes and reports no access to customer files, secrets or application data. Crashes and administrator-level execution were shown only on the researcher's own KVM/Firecracker hosts.
  • Disclosure can happen without a crash. Quiet logs cannot clear historical exposure, and ordinary packet captures may lack the virtual-device metadata.
  • No verified IOCs exist. In-the-wild exploitation was not independently verified.
  • CISA SSVC (16 October 2026) shows exploitation none, automatable no, technical impact total. EPSS is 0.211% (10.655th percentile). The CVE is not in the KEV catalogue of 1,748 entries (2026.10.16).
  • Red Hat's API score of 8.8 is marked draft. NVD shows only the CNA's secondary score.
  • Local exploitability and historical customer impact remain unestablished.

Organisations named here

  • CWE-787
  • PWN.AI
  • PwnEscape

CVEs named here

  • CVE-2026-99999

Technologies named here

  • CISA KEV catalogue
  • CISA SSVC 2.0.3
  • FIRST EPSS
  • Firecracker/KVM
  • NVD
  • Vercel production sandboxes
  • GNU uname
  • AWS bulletin scope for self-managed KVM/virtio-net
  • journalctl
  • Red Hat
  • libvirt
  • Amazon Linux Security Center tracker and ALAS advisories
  • PWN.AI research by Paulos Yibelo
  • Vercel findings blog
  • Linux kernel
  • KVM
  • Firecracker
  • Amazon Linux
  • Red Hat Enterprise Linux
  • Ubuntu
  • Vercel
  • CISA KEV
  • CISA SSVC
  • virtio-net

Standards and laws named here

  • EPSS

Techniques named here

  • Weakness class CWE-787 Out-of-bounds Write
  • Code-path leads for hunting
  • SOC trail C01 to C06
  • Sandbox escape

Also named here

  • Paulos Yibelo