AI Stack Alert Medium CVE-2026-116691 For AppSec, AI Eng, SOC / IR, CISO AI-assisted
MLflow CrewAI autologging writes LLM API keys into trace spans (CVE-2026-116691)
MLflow versions from 2.19.0 before 3.17.0 leak CrewAI LLM provider API keys through mlflow.crewai.autolog(). The autologging serializes the CrewAI Agent's fields, and the llm object's string form contains api_key in plaintext, into span attributes and span inputs. The fix is in MLflow 3.17.0.
Why it matters
Any application using mlflow.crewai.autolog() with an API-keyed CrewAI LLM can send that provider key to the MLflow trace backend or an OTLP destination. Anyone who can read those traces may be able to obtain the key.
Do this
- Upgrade mlflow to 3.17.0 or later in all environments that use mlflow.crewai.autolog().
- Until upgraded, disable mlflow.crewai.autolog() in CrewAI applications.
- Rotate any LLM provider API keys passed to CrewAI LLM objects while autologging was enabled on affected versions.
- Search stored MLflow traces and OTLP destinations for exposed provider keys and restrict read access to them.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.21%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.21%EPSS
- 27/100Priority
- 80%Confidence
Evidence
“an application that calls `mlflow.crewai.autolog()` can unintentionally send its LLM provider API key to the MLflow trace backend or configured OTLP destination.”
“Affected: mlflow >= 2.19.0, < 3.17.0 → fixed in 3.17.0.”
Sources
Affected / related
mlflow, CrewAI, CWE-532 Insertion of Sensitive Information into Log File
Published 2026-10-27 21:26 UTC · updated 2026-11-08 02:39 UTC · AI-assisted summary, checked against quoted sources · How we score