AIKIDO-2026-290660
keras is vulnerable to Unsafe Deserialization
75
CVE Score
High
Unsafe Deserialization
TL;DR
keras.datasets.imdb.load_data and reuters.load_data loaded their .npz archives with np.load(allow_pickle=True), and cifar10/cifar100 batch loading used a bare cPickle.load. A crafted or tampered dataset file can therefore unpickle arbitrary objects and run code through a pickle __reduce__ gadget. The fix routes these loaders through a numpy-only RestrictedUnpickler that permits only array reconstruction, so genuine datasets still load while gadget payloads are refused.
How do I fix this?
Upgrade the keras library to the patch version.
Who does this affect?
You are affected if you are using a version that falls within the vulnerable range and you load IMDB, Reuters, or CIFAR dataset files that an attacker can replace or supply (for example a poisoned local cache or a custom path to untrusted archives).
Package
Versions
3.0.0 - 3.12.3
Fixed in 3.12.43.13.0 - 3.15.0
Fixed in 3.15.1Background info
keras is vulnerable to Unsafe Deserialization in versions 3.0.0 - 3.12.3 and 3.13.0 - 3.15.0.
Get secure today,
quickly and for free.
Secure your code, cloud, and runtime in one central system. Connect a repo to discover what the reasoning agents find in your codebase.
No credit card required | Scan results in 32 seconds.
I consent to receiving marketing communications based on Aikido’s Privacy Policy.
SOC 2Compliant
ISO 27001CompliantISO 42001Compliant
FedRAMPAuthorized© All Intel data is openly available and commercially licensed.