This vulnerability in keras-team/keras version 3.14.0 is due to improper handling of deserialization in the Lambda layer. Specifically, the _raise_for_lambda_deserialization() function conflates a None value for safe_mode (default) with False (explicitly disabled), bypassing the safe-mode guard. This logic flaw allows attacker-controlled marshal bytecode to be deserialized when calling keras.layers.deserialize(config), keras.models.clone_model(model), or Lambda.from_config(config) outside of a SafeModeScope(True) context. Successful exploitation can lead to arbitrary OS-level code execution in the server or user process context. The vulnerability is tracked as CWE-502 (Deserialization of Untrusted Data) and has a CVSS v3.0 score of 8.8 (high severity). The product is a cloud service, and a patch is available though specific patch details are not provided.

Exploitation of this vulnerability can result in arbitrary code execution at the OS level within the context of the server or user process running keras-team/keras. This can lead to full compromise of the affected system. The vulnerability is remotely exploitable without privileges and requires user interaction. The confidentiality, integrity, and availability of the system are all rated high impact.

Since this is a cloud-hosted service, the vendor manages remediation server-side. Users should consult the vendor advisory for confirmation of patch deployment and ensure they are using a version with the fix applied. Avoid invoking Lambda.from_config(config) or related deserialization functions outside of a SafeModeScope(True) context until patched. Monitor vendor communications for updates. Patch status is confirmed as available, but specific patch versions are not listed; check the vendor advisory for current remediation guidance.