Skip to content
AI Sec Watch
MediumVulnerability

CVE-2021-41195: TensorFlow denial of service in tf.math.segment_* operations

Source
NVD(opens in a new tab)
Identifier
CVE-2021-41195
Published
Record updated
View JSON
Affected
  • tensorflow >= 2.6.0, < 2.6.1, fixed in 2.6.1, released November 1, 2021, 5 days before disclosure
  • tensorflow >= 2.5.0, < 2.5.2, fixed in 2.5.2, released November 1, 2021, 5 days before disclosure
  • tensorflow < 2.4.4, fixed in 2.4.4, released November 1, 2021, 5 days before disclosure
  • and 6 more
Known exploitation
Not listed in the CISA Known Exploited Vulnerabilities catalog at the last check.
EPSS
0.2%

Summary

TensorFlow's tf.math.segment_* operations, on both CPU and GPU, abort with a CHECK failure when a segment id in segment_ids is large, causing denial of service. The output shape is computed with AddDim, which fails with a std::abort if the tensor's element count overflows an int64_t. The source attributes the issue to similarity with CVE-2021-29584 and other localized TensorFlow API reports.

Mitigation

The fix will be included in TensorFlow 2.7.0. The commit will also be cherrypicked to TensorFlow 2.6.1, TensorFlow 2.5.2, and TensorFlow 2.4.4, which are affected and still in supported range. The source states that code should use AddDimWithStatus instead of AddDim.

Affected packages in the Exposure Registry

Matched by package name and ecosystem. Each entry shows whether the package delegates to a language model and how many tracked packages depend on it.

Related items