Security1 hOctober 28, 2026, 06:30

MLflow flaw leaks plaintext LLM API keys via CrewAI autologging

CVE-2026-116691 (CVSS 6.5) in MLflow 2.19.0 through 3.16.x exposes plaintext LLM provider API keys in OpenTelemetry trace attributes when mlflow.crewai.autolog() is enabled. Keys are sent to MLflow tracking servers or OTLP backends. The flaw is fixed in MLflow 3.17.0, and affected credentials should be rotated.

DEV Community

Open source

MLflow flaw leaks plaintext LLM API keys via CrewAI autologging

Affects MLflow 2.19.0 to 3.16.x with mlflow.crewai.autolog()
CVSS 6.5, CWE-532: secrets written to logs and traces
Fixed in MLflow 3.17.0; rotate exposed API keys
Public proof-of-concept exists; not in CISA KEV

#MLflow#CrewAI

MLflow RCE CVE-2026-116692 rated 9.8 via scorer deserialization

Security

MLflow flaw allows arbitrary file read via prompt model versions

Security

MLflow CVE-2026-64849: unauthenticated SSRF via webhook endpoint

Security

MLflow flaw leaks plaintext LLM API keys via CrewAI autologging

MLflow RCE CVE-2026-116692 rated 9.8 via scorer deserialization

MLflow flaw allows arbitrary file read via prompt model versions

MLflow CVE-2026-64849: unauthenticated SSRF via webhook endpoint

METR breach: LLM key abuse burned $600,000 in three weeks