Yao has native MCP support and TS Create/Next hooks that intercept every request — a clean pre-invoke point. There's no threat-detection layer on tool I/O yet.
ATR (Agent Threat Rules, MIT) is an open detection-rule standard for agent attacks (prompt injection, tool poisoning, context/credential exfil, MCP attacks), pure rule eval (no model call). I can PR a small yao-atr hook module that runs an ATR scan inside the Next hook on each MCP tool invocation and blocks/annotates on match. Would you take it as an optional module?
Yao has native MCP support and TS Create/Next hooks that intercept every request — a clean pre-invoke point. There's no threat-detection layer on tool I/O yet.
ATR (Agent Threat Rules, MIT) is an open detection-rule standard for agent attacks (prompt injection, tool poisoning, context/credential exfil, MCP attacks), pure rule eval (no model call). I can PR a small yao-atr hook module that runs an ATR scan inside the Next hook on each MCP tool invocation and blocks/annotates on match. Would you take it as an optional module?