Autonomous AI agents promise super-charged productivity but without the right guardrails they can also jailbreak, leak data, or go off-topic. In this session we will discuss about:

• building a lightweight agentic workflow from scratch
• probing real-world vulnerabilities, AI risks
• mapping your agent against the four-axis “agentic profile” framework for alignment and governance
• applying risk-mitigation checklists distilled from risk management frameworks

What we will build -
In the hands-on segment we will build a complete agent to go from blank notebook to governed production prototype. We’ll begin by bootstrapping a one-file Python agent with LangChain and OpenAI Functions that can plan, call external APIs, and write concise summaries. Next, we’ll wrap that agent with the open-source Python libraries, layering in rate-limits, PII scrubbing, and role-based tool permissions so you can see policy enforcement in action. With guardrails in place, we’ll shift to offense - running an automated PyTest suite populated with the red-team prompts to expose prompt-injection and tool-abuse weak spots. We’ll then quantify how well the patched agent stays on-mission by applying a lightweight PRISM-style alignment rubric that emits a JSON scorecard. Finally, we’ll wire everything into a Streamlit mini-dashboard that streams agent actions, policy hits, and manual override controls in real time, giving a turnkey template we can fork for our next project.