Junglewise Threat Intelligence

CVE-2026-2492: TensorFlow privilege escalation via uncontrolled HDF5 plugin search path

CVE-2026-2492 · Severity: high · CVSS 7.8 · Published 2026-02-20

Technologies: Google TensorFlow. Vendors: Google, Red Hat.

Executive brief

A vulnerability in the TensorFlow machine learning library could allow a user with limited access to a system to gain higher-level permissions. This occurs because the software attempts to load plugin components from an insecure folder location that a malicious user can control. If exploited, an attacker could execute their own code with the privileges of another user, potentially compromising sensitive data or system integrity.

Technical details

A flaw exists in TensorFlow's handling of HDF5 library plugins (CWE-427). The application attempts to load plugins from an unsecured or uncontrolled search path, allowing a local attacker to place a malicious library in the path. To exploit this, an attacker must already have the ability to execute low-privileged code on the target system. Successful exploitation leads to arbitrary code execution and privilege escalation to the level of the user running the TensorFlow application. The fix involves disabling external HDF5 plugin loading by setting the HDF5_PLUGIN_PATH environment variable to 'disable' during the h5py import process.

Affected products

  • TensorFlow TensorFlow 2.17.0
  • Red Hat Red Hat OpenShift AI 2.25

Timeline

  • 2025-03-11: disclosed: Vulnerability reported to vendor
  • 2026-02-19: advisory: ZDI advisory published
  • 2026-02-20: disclosed: CVE published to NVD
  • 2026-04-23: patched: Red Hat released updates for OpenShift AI 2.25.5

References

Related threats