safety-alignment

Overview

A collection of 4 AI safety and alignment research skills from Orchestra Research, providing engineering ability for coding agents to conduct AI safety experiments. Part of a larger collection of 83 AI research engineering skills covering model architecture, fine-tuning, tokenization, evaluation, agents, RAG, multimodal, and more.

Key Features

• Constitutional AI — AI-driven self-improvement via principles (282 lines of structured instructions)
• LlamaGuard — Safety classifier for LLM inputs/outputs (329 lines)
• NeMo Guardrails — Programmable guardrails with Colang for controlling LLM behavior (289 lines)
• Prompt Guard — Meta's 86M parameter prompt injection and jailbreak detector with 99%+ TPR and under 2ms GPU latency (313 lines)
• Research-oriented — Each skill includes reference materials and structured instructions for hands-on experimentation
• Part of 83 skills — Spans 15 categories from model architecture to MLOps to mechanistic interpretability

Who is this for?

AI safety researchers and ML engineers who want their coding agent to help with safety alignment experiments — from training guardrails to evaluating prompt injection defenses. Ideal for teams working on responsible AI deployment who need structured, reproducible safety engineering workflows.