# CVE-2026-96775: MLflow dspy flavor insufficient pickle deserialization control

Source: https://junglewise.ai/threats/cve-2026-96775-mlflow-s-dspy-flavor-versions-2-0-applies-the-mlflow-allow-pickle  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-09-23. Updated: 2026-09-26  
License: CC BY 4.0 with a link back

CVE-2026-96775 · Severity: high · CVSS 8.8 · Published 2026-09-23

MLflow is a platform for managing machine learning models and their deployment pipelines. The dspy flavor used to load models has a security control that can be bypassed by naming a pickle file with an extension other than .pkl, allowing an attacker who can write to a model repository to execute arbitrary code when a user loads the model.

Technologies: [Mlflow](https://junglewise.ai/threats/technologies/mlflow-mlflow). Vendors: [MLflow](https://junglewise.ai/threats/vendors/mlflow).

## Quick answers

**Is CVE-2026-96775 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-96775 being exploited in the wild.

**How severe is CVE-2026-96775?** High severity, with a CVSS score of 8.8, published 23 September 2026.

**Which products and versions are affected?** Mlflow >= 2.0.

**Is there a fix for CVE-2026-96775?** No fixed version is recorded yet. Check the vendor's advisory in the references below. If you run it, Junglewise drafts a mitigation that fits your setup and applies it once you approve, with senior security engineers behind it.

**How is CVE-2026-96775 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-96775?** Check whether you run Mlflow at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## Technical details

The vulnerability exists in MLflow's dspy flavor, which fails to apply the MLFLOW_ALLOW_PICKLE_DESERIALIZATION safety control when the model path does not end in .pkl, despite the file being a pickle file. An attacker with write access to a model storage location can craft a malicious MLmodel artifact with a non-.pkl extension to bypass the security check and achieve arbitrary code execution via pickle deserialization. A fix for the statsmodels flavor was released in versions >= 3.15.0, but the dspy flavor remains unpatched.

## Affected products

- MLflow MLflow >= 2.0

## Timeline

- 2026-09-16: disclosed
- 2026-09-23: advisory

## References

- [https://kb.cert.org/vuls/id/369093](https://kb.cert.org/vuls/id/369093)

## Related threats

- [MLflow path traversal in FileStore log outputs API](https://junglewise.ai/threats/cve-2026-127202-mlflow-before-3-18-0-using-the-filestore-backend-contains-a-path)
- [MLflow get-artifacts server-side request forgery information disclosure](https://junglewise.ai/threats/cve-2026-118628-mlflow-get-artifacts-server-side-request-forgery-information)
- [MLflow SHAP load_explainer deserialization of untrusted data](https://junglewise.ai/threats/cve-2026-118627-mlflow-shap-load-explainer-deserialization-of-untrusted-data)
- [MLflow Tracking Server missing authorization in get_endpoints](https://junglewise.ai/threats/cve-2026-118625-mlflow-tracking-server-missing-authorization-privilege)
- [MLflow Tracking Server authorization bypass in Jobs API](https://junglewise.ai/threats/cve-2026-118624-mlflow-tracking-server-jobs-missing-authorization-information)