CVE-2025-9906
Keras is vulnerable to Deserialization of Untrusted Data
Impact
### Arbitrary Code Execution in Keras Keras versions prior to 3.11.0 allow for arbitrary code execution when loading a crafted `.keras` model archive, even when `safe_mode=True`. The issue arises because the archive’s `config.json` is parsed before layer deserialization. This can invoke `keras.config.enable_unsafe_deserialization()`, effectively disabling safe mode from within the loading process itself. An attacker can place this call first in the archive and then include a `Lambda` layer whose function is deserialized from a pickle, leading to the execution of attacker-controlled Python code as soon as a victim loads the model file. Exploitation requires a user to open an untrusted model; no additional privileges are needed. The fix in version 3.11.0 enforces safe-mode semantics *before* reading any user-controlled configuration and prevents the toggling of unsafe deserialization via the config file. **Affected versions:** < 3.11.0 **Patched version:** 3.11.0 It is recommended to upgrade to version 3.11.0 or later and to avoid opening untrusted model files.
Affected packages
| Ecosystem | Package | Affected | Fixed in |
|---|---|---|---|
| pip | keras | < 3.11.0 | 3.11.0 |
Affected products
| Vendor | Product | Affected versions |
|---|---|---|
| Keras | Keras | from 3.0.0 before 3.11.0 |
Weakness classification
References
- NVD record
- CVE.org record
- GitHub advisory
- https://github.com/keras-team/keras/pull/21429
- https://github.com/advisories/GHSA-36fq-jgmw-4r9c
- https://nvd.nist.gov/vuln/detail/CVE-2025-9906
- https://github.com/keras-team/keras/commit/713172ab56b864e59e2aa79b1a51b0e728bba858
- https://github.com/keras-team/keras/releases/tag/v3.11.0
- https://osv.dev/vulnerability/CVE-2025-9906
- https://github.com/pypa/advisory-database/tree/main/vulns/keras/PYSEC-2025-76.yaml