Autonomous security reasoning at Microsoft Hyderabad

I'm Arjun Trivedi, a security researcher at Microsoft in Hyderabad. I build AI tooling for security investigations. The main project is ASRS (Autonomous Security Reasoning System), which fuses alerts across multiple SIEMs, normalizes them to OCSF, and lets you query them in natural language. It has observe, assist and autonomous modes, and in assist mode you approve every action. I also maintain two open source MCP servers: fastmcp-threatintel for threat intelligence and mcp-kql-server for running KQL from an AI prompt.

My day-to-day stack is LangGraph with Claude, Gemini, Qwen via Ollama, plus Codex and Copilot. I'm still working out how much autonomy is appropriate for investigations, and where a human should stay in the loop.

I'd love feedback from anyone who has put agentic systems into high-stakes workflows: how did you design approval and trust boundaries?