Allocation of Resources Without Limits or Throttling in Keras - CVE-2026-0897
Published: November 9, 2026
Vulnerability details
The vulnerability allows an unauthenticated remote attacker to cause memory exhaustion and denial of service in environments processing untrusted Keras models.
The vulnerability exists due to unrestricted memory allocation from unvalidated HDF5 dataset dimensions when KerasFileEditor loads a crafted .keras archive. An unauthenticated remote attacker can supply a small model file declaring enormous dataset dimensions and persuade a victim to open it, triggering a petabyte-scale allocation attempt that crashes the process.