Technical impact
MLflow versions 2.19.0 through 3.17.0 allow the CrewAI autologging feature to capture agent objects without sanitizing nested credential information. The internal conversion of the llm field to a string includes the provider API key, and the expansion of the agent object into span inputs can send the key to the trace backend or OTLP exporter. As a result, plaintext provider credentials become visible in trace logs, presenting a clear confidentiality breach that could expose sensitive API keys. The vulnerability does not grant code execution but can compromise any secrets that the application stores for language‑model providers.
Affected Systems
The affected product is MLflow. The vulnerability exists in all releases from 2.19.0 up to and including 3.17.0, with the fix applied in the 3.17.0 release and later. Users on earlier or later versions are not impacted.
Risk and Exploitability
The CVSS score of 6.5 indicates a medium severity level. Exposure is possible through any application that invokes CrewAI autologging, and the likely attack vector, inferred from the description, is local to the code execution environment. While official exploitation evidence is not reported and the EPSS score of 0.00207 (approximately 0.21%) indicates a very low exploitation probability, the absence of this vulnerability from CISA’s KEV catalog suggests it is not currently widely exploited. Nonetheless, the potential for leaking sensitive API keys makes this a significant risk to confidentiality, especially when trace data is forwarded to external OTLP services or stored in accessible locations.
Github GHSA