CVE-2023-52305 is a medium severity (4.7) division by zero crash in paddle.topk in PaddlePaddle before 2.6.0. Bad arguments can stop the process and cause a denial of service. The fix ships in 2.6.0, so upgrade PaddlePaddle to 2.6.0 or later and validate untrusted input.
What's affected
The NVD record names PaddlePaddle, the open source deep learning framework from the PaddlePaddle project, in every release before 2.6.0. The project's own advisory, PDSA-2023-014, says the fix will be included in PaddlePaddle 2.6.0 and points to the commit 19da5c0c4d8c.
The weakness classes are CWE-369. The record's vector is CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:C/C:N/I:N/A:L, which gives 4.7 (Medium). The advisory itself does not state a score.
The advisory titles the problem as a floating point exception (division by zero) in paddle.topk. A floating point exception in this context is a division by zero in native code, reached because a size or parameter was not checked first. The process stops. The weakness class is CWE-369, and the effect is a crash, so the record rates availability impact only.
How to check if you're exposed
Paddle ships on PyPI as the paddlepaddle package, and GPU builds use their own package names, so search for anything containing paddle. In the Python environment where you run the framework:
pip list | grep -i paddle
python -c "import paddle; print(paddle.__version__)"
Any version below 2.6.0 is in the affected range. Check every environment, including virtual environments, notebooks, training servers, containers and CI images, since each one carries its own copy.
To find where the library is pulled in, search your requirements files, pyproject files and lockfiles for paddle. Then ask a second question: does any code path hand untrusted input to paddle.topk? The advisory describes the weakness at the level of that one function, so your real exposure depends on whether your application reaches it.
What to do now
- Find every environment that has PaddlePaddle installed, using the commands above.
- Upgrade to 2.6.0 or later, for example with
pip install -U paddlepaddle, then run your tests, since a framework release can change behaviour. - Pin the new version in your requirements file or lockfile so a rebuild does not bring the old one back.
- Until you upgrade, do not pass untrusted values to paddle.topk. Validate shapes, sizes and parameters before they reach the framework.
- Rebuild and redeploy container images that bundle the framework, because old layers keep the old library.
Why it matters
A floating point exception in this context is a division by zero in native code, reached because a size or parameter was not checked first. The process stops. The weakness class is CWE-369, and the effect is a crash, so the record rates availability impact only.
Most applications never feed attacker controlled arguments into this operation. A model served over an API, or a service that accepts user supplied parameters, is where it matters. This CVE is one of a group of PaddlePaddle issues disclosed together and fixed in the same release, so the single upgrade covers its siblings. It is not in the CISA Known Exploited Vulnerabilities catalog, and the advisory does not report exploitation in the wild. Safeguard's reachability analysis is built for this question, whether a vulnerable function is actually called by your code. You can also run the free in-browser dependency scan at gold.safeguard.sh/scan.
Quick check
Which PaddlePaddle versions does CVE-2023-52305 affect?
The NVD record and the advisory PDSA-2023-014 say versions before 2.6.0. The fix is included in 2.6.0, so that release or later is the target.
Does CVE-2023-52305 affect my application if I never call paddle.topk?
The advisory describes the weakness in that one function, so code that never reaches it is not exposed through this CVE. Confirm by searching your code and your dependencies, and upgrade anyway because the fix is a routine release.
Sources
Last reviewed: 19 October 2026.
