CVE-2023-33976 is a high severity (CVSS 7.5) denial of service in TensorFlow: array_ops.upper_bound causes a segmentation fault when it is not given a rank 2 tensor. TensorFlow says the fix ships in 2.13 and is cherry-picked to 2.12. Upgrade to 2.13.0 or later.
What's affected
- TensorFlow, tensorflow-cpu, and tensorflow-gpu on PyPI, versions before 2.13.0, per the GitHub advisory.
- The advisory lists patched version 2.12.0, and its text says the fix lands in 2.13 and is cherry-picked to 2.12.0. Use 2.13.0 or later if you can.
- The advisory links two fix commits in the TensorFlow repository. The NVD record lists CWE-190, integer overflow or wraparound. The GitHub advisory lists no CWE.
How to check if you're exposed
Check the installed version in each environment:
pip show tensorflow tensorflow-cpu tensorflow-gpu
Or print it from Python:
python -c "import tensorflow as tf; print(tf.__version__)"
Also search your lockfiles and requirements files for the three package names, because model serving images and notebooks often pin an older release. Container images need the same check inside the image, not on the host.
Then ask whether any code path calls the upper bound operation with data from outside your trust boundary. If shapes are fixed by your own code, the risk is lower.
What to do now
- Upgrade tensorflow, tensorflow-cpu, or tensorflow-gpu to 2.13.0 or later, and run your model tests.
- Rebuild and redeploy container images that bundle TensorFlow so the new version is what runs.
- Validate tensor rank and shape on inputs that come from users or external systems before calling the operation.
- Run model serving behind restarts and health checks so a crash does not become an outage.
- Add TensorFlow to your dependency monitoring so the next advisory finds every environment.
If you track software dependencies, the free in-browser dependency scan can show which of your components carry known CVEs, and the Gold directory lists the same records in one place.
Why it matters
A segmentation fault is a process crash. For a service that serves models, a crash on malformed input is a denial of service. The advisory says the operation fails when it receives something other than a rank 2 tensor, which is an input validation gap inside the library.
The disagreement about severity is worth noting. The NVD record scores it High at 7.5. The GitHub advisory rates it Moderate. Both are fair views: the network vector looks serious on paper, but the effect is a crash and only matters where outside input reaches the operation.
The record shows this CVE is not in the CISA Known Exploited Vulnerabilities catalog.
Try it yourself
Click through it at your own pace, or open it full screen.
Quick check
Why do the sources disagree on severity?
NVD scores CVSS 7.5 High. The GitHub advisory labels it Moderate. Use High to prioritize patching and Moderate as a hint that the impact is a crash rather than data loss.
Which version should I use?
The advisory says the fix is in TensorFlow 2.13 and cherry-picked to 2.12.0. Pick 2.13.0 or later unless your platform is bound to 2.12.
Sources
Last reviewed: 19 October 2026.