Cloudanix VulnDB

CVE-2026-98143

High · 7.8

CVSS

7.8

EPSS

0.1%

Published

Sep 25, 2026

CISA due date

—

Cloudanix analysis

Next patch cycle SSVC: Track

A driver bug in the Linux kernel can make certain artificial-intelligence calculations on specific chips either crash or give wrong answers.

Exploitation
Local Privilege
Attacker needs
Low Privilege
Complexity
Low
Impact
Full Compromise
Component
OS kernel
Class
Improper Input Validation
Remediation
Patch
Effort
Medium
Patch risk
Low

What it actually does

The ethosu driver misinterprets bits 15:14 of NPU_SET_{IFM,OFM}_PRECISION as storage mode on U65 hardware where they instead encode rounding mode, causing either -EINVAL rejection of valid command streams or selection of an incorrect two-tile address path that derives bogus feature map sizes.

Why it matters to the business

Machine-learning workloads using average pooling, concatenation, split, unpack, strided slice, LUT or argmax on U65-based accelerators fail or produce silently incorrect results, breaking inference pipelines and test suites on platforms such as i.MX93.

How an attack unfolds

A local attacker submits a command stream containing natural rounding (2) or truncating rounding (1) values; the parser either rejects the buffer creation or computes wrong tile addresses, resulting in denial of service or corrupted accelerator output.

How to fix it

Upgrade the Linux kernel to a version containing the commit that conditions storage-mode interpretation on U85 hardware only.

How to confirm the fix worked

Confirm the driver source no longer reads bits 15:14 as storage mode on U65; re-run the Teflon test suite covering the seven affected operations and verify absence of CMDSTREAM_BO_CREATE -EINVAL errors.

Typical attack path

  1. Obtain local access on a system with the ethosu driver loaded
  2. Generate a command stream using Mesa or equivalent that requests natural or truncating rounding
  3. Invoke CMDSTREAM_BO_CREATE via the accelerator ioctl path
  4. Trigger either -EINVAL failure or miscomputed feature-map sizes

How to detect it

Search kernel logs and dmesg for ethosu CMDSTREAM_BO_CREATE failures returning -EINVAL, especially on systems containing U65 hardware; correlate with processes invoking Mesa average-pooling or argmax operations.

  • CMDSTREAM_BO_CREATE.*-EINVAL
  • ethosu.*rounding
  • U65.*precision

What your security team should do

  • Prioritize kernel updates on all systems exposing the ethosu driver
  • Add monitoring for ethosu-related -EINVAL errors in kernel logs
  • Restrict unprivileged access to the accelerator device node until patched

The record

In the Linux kernel, the following vulnerability has been resolved: accel: ethosu: Don't read the U65 rounding mode as a storage mode Bits 15:14 of NPU_SET_{IFM,OFM}_PRECISION select the activation storage mode on U85 only. On U65 the same field holds the rounding mode, and the command stream parser has read it as a storage mode since the driver was added. That went unnoticed while unknown values fell through the switch, but now that they are rejected, every U65 command stream that asks for natural rounding (2) fails CMDSTREAM_BO_CREATE with -EINVAL. Mesa emits it for average pooling, concatenation, split, unpack, strided slice, LUT and argmax, which is 72 failures of the Teflon test suite on an i.MX93. Truncating rounding (1) is misread as well: it picks the two-tile address path and computes a bogus feature map size from tile bases the command stream never set. Read the field as a storage mode only on the hardware where it is one.

Affected packages

Ecosystem Package Introduced Fixed