Description
The Keras Model.load_model method can be exploited to achieve arbitrary code execution, even with safe_mode=True.
One can create a specially crafted .keras model archive that, when loaded via Model.load_model, will trigger arbitrary code to be executed. This is achieved by crafting a special config.json (a file within the .keras archive) that will invoke keras.config.enable_unsafe_deserialization() to disable safe mode. Once safe mode is disable, one can use the Lambda layer feature of keras, which allows arbitrary Python code in the form of pickled code. Both can appear in the same archive. Simply the keras.config.enable_unsafe_deserialization() needs to appear first in the archive and the Lambda with arbitrary code needs to be second.
Affected products
| Vendor | Product | Affected | Fixed |
|---|---|---|---|
| Keras-team | Keras | 3.0.0 – < 3.11.0 | 3.11.0 |
Fixed in Linux distributions
Distributions backport fixes, so the version that fixes this vulnerability differs per release — upgrade to at least the version shown. From the distributions' security trackers and advisories.
Debian
Security tracker ↗| Release | Package | Fixed in | Advisory |
|---|---|---|---|
| Debian 11 | keras | No fix yet | — |
Ubuntu
Security tracker ↗| Release | Package | Fixed in | Advisory |
|---|---|---|---|
| Ubuntu 20.04 LTS | keras | No fix yet | — |
| Ubuntu 18.04 LTS | keras | No fix yet | — |
Azure Linux
| Release | Package | Fixed in | Advisory |
|---|---|---|---|
| Azure Linux 3 | keras | 3.3.3-4 | — |
Wolfi
| Release | Package | Fixed in | Advisory |
|---|---|---|---|
| Wolfi rolling | kubeflow-pipelines-visualization-server | 2.17.0-r0 | — |
Affected packages
| Ecosystem | Package | Affected | Fixed | Record |
|---|---|---|---|---|
| PyPI | keras | < 3.11.0 | 3.11.0 | GHSA-36fq-jgmw-4r9c |
| PyPI | keras | — | — | PYSEC-2025-76 |