I have recently reported CVE-2026-116691, that I found in MLFlow - one of the most popular platform for ML (20M+ Weekly downloads, used by DataBricks, Walmart, Spotify).
I found the API as Plain text.
How?
The instrumentation layer - MLFlow had did the serialization of the LLM's Pythonic object and placed it in plain text into its trace.
Something that went over all of the cummon Security best practices - the Least Priviledge Access model.
You can use identity cotrols access. How can you control data flow?
Thats why I've built Weir - it can be used to making sure what AI Agents has actual access to - as SDK with no egress what so ever.
What do you think? how can you guard your agent and your agent's harness agents such leaks?
1 comment
[ 3.2 ms ] story [ 30.0 ms ] threadI found the API as Plain text.
How?
The instrumentation layer - MLFlow had did the serialization of the LLM's Pythonic object and placed it in plain text into its trace.
Something that went over all of the cummon Security best practices - the Least Priviledge Access model.
You can use identity cotrols access. How can you control data flow?
Thats why I've built Weir - it can be used to making sure what AI Agents has actual access to - as SDK with no egress what so ever.
What do you think? how can you guard your agent and your agent's harness agents such leaks?