MLflow wrote LLM API keys in plain text to its logs (CVE-2026-116691)
TL;DR Agent tracing (CrewAI autologging) in MLflow 2.19.0 to 3.16.x recorded API keys in clear text. Fixed in 3.17.0.
In plain words: MLflow is a popular tool to record what machine-learning and AI-agent runs did. When it recorded agent runs automatically, it also wrote down the secret keys used to call the AI providers, readable by anyone with access to those records.
What happened
In MLflow 2.19.0 through 3.16.x, automatic agent tracing (CrewAI autologging) stored the LLM provider API keys in plain text in logs and traces.
Who is affected
Teams that used MLflow autologging for agents in those versions, plus anyone who can read the tracking server or wherever its logs were copied.
Why it matters
Logs travel: they get shipped to storage buckets, shared in tickets, attached to bug reports. A key in a log ends up in many more places than the key itself.
What to do
- Upgrade to 3.17.0.
- Rotate the keys that went through those logs.
- Check where the logs were stored and copied.
- In general: any agent log or transcript can contain secrets. Scan them periodically.