Become an AI engineer
Safety, plainly
Safety, plainly is a unit of 16 lessons in Rabbithole's Become an AI engineer course, at levels 1 to 2.
Lessons
- AI safety: what it is and how people think about itLevel 1
- Alignment: an AI that does what we meanLevel 2
- Jailbreaks: Getting an AI to Break Its Own RulesLevel 2
- Red-teaming: attacking an AI on purpose to find its flawsLevel 2
- Safety frameworks: how AI companies decide to launchLevel 2
- Testing AI for dangerous abilitiesLevel 2
- Model Weight SecurityLevel 2
- Model welfare: could an AI have wellbeing?Level 2
- Reward hacking: when an AI games its scoreLevel 2
- Agentic MisalignmentLevel 2
- Alignment faking: when an AI plays alongLevel 2
- Corrigibility: not working against human oversightLevel 2
- Data Poisoning: Bad Data In, Bad Behavior OutLevel 2
- Deliberative alignment: a model that reasons over written safety rulesLevel 2
- Sandbagging: when an AI holds back on a testLevel 2
- Scalable oversight: keeping AI answers checkableLevel 2
Sources
These lessons draw on 29 sources. The most used:
- Anthropic Risk Report, August 2026
- GPT-4 Technical Report
- Alignment faking in large language models
- Deliberative Alignment: Reasoning Enables Safer Language Models
- Frontier Safety Framework Report — Gemini 3 Pro
- Sabotage evaluations for frontier models
- Claude's Constitution
- Our approach to alignment research