GPU VulnDB

Database/AI/ML frameworks & serving

Hugging Face Transformers: convert_config executes strings from a checkpoint in three speech model converters

CVSS 7.8CVE-2025-14926AI/ML frameworks & serving+2 more CVEscurated

Impact

The convert_config function in the SEW, SEW-D and HuBERT conversion scripts takes a string out of the checkpoint and uses it to execute Python. Converting an untrusted checkpoint therefore runs attacker code as the user doing the conversion, with that user's access to model stores, cluster credentials and GPU devices. ZDI split this across three ids - CVE-2025-14926 (SEW, ZDI-25-1147), CVE-2025-14927 (SEW-D, ZDI-25-1148) and CVE-2025-14928 (HuBERT, ZDI-25-1146) - all rated 7.8 against Transformers 4.57.0, all the same injection in the same helper, with the same operator response for all three.

Who can reach it

No authentication against the library. Someone has to run a conversion script on an attacker-supplied checkpoint - typically an ingest pipeline or an engineer onboarding third-party speech weights. Code runs as that user.

What to do

No fixed version is named in the record for any of the three ids. Treat the conversion scripts as code execution on their input: run them in a disposable container, as a low-privilege user, with no cluster credentials and read-only access to shared model storage. A fixed Transformers release is a pip upgrade; nothing at the node level.

Also covers 2 CVEs

The vendor assigned a separate id to each affected code path. They share this advisory, this score and this fix, so they are one entry here.

CVE-2025-14927CVE-2025-14928

References

https://nvd.nist.gov/vuln/detail/CVE-2025-14926https://www.zerodayinitiative.com/advisories/ZDI-25-1147/https://www.zerodayinitiative.com/advisories/ZDI-25-1148/https://www.zerodayinitiative.com/advisories/ZDI-25-1146/

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.