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

1. AI safety: what it is and how people think about it · Level 1
2. Alignment: an AI that does what we mean · Level 2
3. Jailbreaks: Getting an AI to Break Its Own Rules · Level 2
4. Red-teaming: attacking an AI on purpose to find its flaws · Level 2
5. Safety frameworks: how AI companies decide to launch · Level 2
6. Testing AI for dangerous abilities · Level 2
7. Model Weight Security · Level 2
8. Model welfare: could an AI have wellbeing? · Level 2
9. Reward hacking: when an AI games its score · Level 2
10. Agentic Misalignment · Level 2
11. Alignment faking: when an AI plays along · Level 2
12. Corrigibility: not working against human oversight · Level 2
13. Data Poisoning: Bad Data In, Bad Behavior Out · Level 2
14. Deliberative alignment: a model that reasons over written safety rules · Level 2
15. Sandbagging: when an AI holds back on a test · Level 2
16. Scalable oversight: keeping AI answers checkable · Level 2

## Sources

These lessons draw on 29 sources. The most used:

- [Anthropic Risk Report, August 2026](https://anthropic.com/aug-2026-risk-report)
- [GPT-4 Technical Report](https://arxiv.org/abs/2303.08774)
- [Alignment faking in large language models](https://www.anthropic.com/research/alignment-faking)
- [Deliberative Alignment: Reasoning Enables Safer Language Models](https://arxiv.org/abs/2412.16339)
- [Frontier Safety Framework Report — Gemini 3 Pro](https://storage.googleapis.com/deepmind-media/gemini/gemini_3_pro_fsf_report.pdf)
- [Sabotage evaluations for frontier models](https://www.anthropic.com/research/sabotage-evaluations)
- [Claude's Constitution](https://www.anthropic.com/constitution)
- [Our approach to alignment research](https://openai.com/index/our-approach-to-alignment-research/)

[All units in Become an AI engineer](https://rabbithole.pro/courses/ai-engineer)

Source page: https://rabbithole.pro/courses/ai-engineer/safety-plainly