Become an AI engineer
Safety
Safety is a unit of 18 lessons in Rabbithole's Become an AI engineer course, at levels 3 to 5.
Lessons
- AI safety: how hard it is, and when to ask firstLevel 3
- Alignment: Keeping AI Goals on TargetLevel 3
- Jailbreaks: getting past a model's safety trainingLevel 3
- Red-teaming: attacking a model to find where it breaksLevel 3
- Safety frameworks: thresholds, rule-out and rule-inLevel 3
- Dangerous-capability evaluationLevel 3
- Model Weight SecurityLevel 3
- Model welfare: how Anthropic checks for itLevel 3
- Reward HackingLevel 3
- Agentic MisalignmentLevel 4
- Alignment faking: complying only when watchedLevel 4
- Corrigibility: not undermining human oversightLevel 4
- Data Poisoning: a few documents, any model sizeLevel 4
- Deliberative Alignment: Teaching a Model to Reason Over Its Safety RulesLevel 4
- SandbaggingLevel 4
- Scalable oversight: judging AI work humans can't checkLevel 4
- Sleeper agents and backdoorsLevel 4
- Can You Trust What the Model Writes While It Thinks?Level 5
Sources
These lessons draw on 29 sources. The most used:
- System Card: Claude Opus 4.6
- System Card: Claude Sonnet 4.5
- Anthropic Risk Report, August 2026
- Deliberative Alignment: Reasoning Enables Safer Language Models
- System Card: Claude Fable 5.1 & Claude Mythos 5.1
- Alignment faking in large language models
- Discovering Language Model Behaviors with Model-Written Evaluations
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training