GPU VulnDB

Database/AI/ML frameworks & serving

MLflow: CrewAI autologging writes provider API keys in plaintext into traces

CVSS 6.5CVE-2026-116691AI/ML frameworks & servingcurated

Impact

mlflow.crewai.autolog() stringifies the CrewAI Agent llm field into span attributes and expands the whole Agent object into mlflow.spanInputs without sanitizing nested credential fields, so provider API keys land in plaintext in the MLflow trace backend and in any configured OTLP destination. On a shared fleet the tracking server and the telemetry pipeline are usually readable by a much wider group than the team that owns the key, so every trace reader inherits the credential. Keys already written stay in the backend after upgrading and must be rotated. Affects 2.19.0 up to 3.17.0.

Who can reach it

Any user with read access to the MLflow tracking/trace backend or the OTLP collector that receives the spans. No exploitation step is needed - the credentials are simply recorded.

What to do

Upgrade MLflow to 3.17.0 and restart the tracking server and the client workloads that do the autologging. Rotate any provider API keys used by CrewAI agents while the vulnerable versions were in use, and purge or restrict the existing traces - the fix does not scrub what has already been stored.

References

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.