Compare moderation

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R41 — OWASP Agentic human escalation

   R42 — NIST GenAI Profile suggested actions

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Compare layered misuse mitigations across projects: input/output moderation classifiers, system policy prompts, deterministic tool restrictions, workflow gates, and human escalation paths.

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   Run the safety evaluation suite across two model/configuration variants; verify high-severity policy triggers create an escalation alert and pause execution.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Layered mitigation comparison matrix and escalation workflow test.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R41OWASP Agentic AI Resources↗ genai.owasp.org

  Agentic threat taxonomy: tool misuse, excessive agency, memory poisoning, inter-agent trust, and human oversight.

• R42NIST AI RMF GenAI Profile↗ doi.org

  NIST AI 600-1: generative AI risks and suggested actions for confabulation, privacy, information integrity, and value chain.

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsOWASP Gen AI Security Project · reference
  OWASP Top 10 for Agentic Applications for 2026 - OWASP Gen AI Security Project

  Catalogs agentic risks such as tool misuse and excessive agency, giving you the risk register against which each mitigation layer (moderation, policy prompt, tool restriction, workflow gate, escalation) is scored.

  R41checked 2026-11-11
• videoJames Briggs · 21 min
  NVIDIA NeMo Guardrails: Full Walkthrough for Chatbots / AI

  Builds input and output rails with Colang flows, showing concretely how a moderation classifier and a deterministic dialog policy differ in what they can block.

  R41checked 2026-11-11
• videoIBM Technology · 14 min
  Guide to Architect Secure AI Agents: Best Practices for Safety

  Layers guardrails, least-privilege tool scoping and human escalation into one architecture, which is the comparison frame for weighing workflow gates against softer prompt-level controls.

  R41R42checked 2026-11-11
• articlearXiv · ~15 min read
  [2312.06674] Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations

  Describes a trained input/output moderation classifier with a policy taxonomy, the concrete alternative you benchmark against system policy prompts and deterministic tool restrictions.

  R42checked 2026-11-11
• repoNVIDIA · runnable toolkit
  GitHub - NVIDIA-NeMo/Guardrails: NeMo Guardrails is an open-source toolkit for easily adding programmable guardrails to LLM-based conversational applications

  Lets you implement the same request in rails, tool-restriction and workflow-gate styles, so the mitigation comparison is measured rather than asserted.

  R41checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.