Curious how others are thinking about approaching agentic security for MCP connected agents. External tools processing raw data from websites or other APIs with no sanitization.
There is no distinction between system instructions and whatever the tool pulled back. Wondering if there is an existing tool or service that fills that gap.
Are you doing anything to monitor or filter tool response traffic before it hits the model? Running in isolated environments? Treating all tool output as untrusted?
Building something in this space and trying to understand how the community is approaching this (if at all?)
Are you doing anything to monitor or filter tool response traffic before it hits the model? Running in isolated environments? Treating all tool output as untrusted?
Building something in this space and trying to understand how the community is approaching this (if at all?)