← Back to curriculum

Phase 18 · 30 lessons

Ethics, Safety & Alignment

Build AI that helps humanity. Not optional.

01Instruction-Following as Alignment SignalLearn 02Reward Hacking and Goodhart's LawLearn 03The Direct Preference Optimization FamilyLearn 04Sycophancy as RLHF AmplificationLearn 05Constitutional AI and RLAIFLearn 06Mesa-Optimization and Deceptive AlignmentLearn 07Sleeper Agents — Persistent DeceptionLearn 08In-Context Scheming in Frontier ModelsLearn 09Alignment FakingLearn 10AI Control — Safety Despite SubversionLearn 11Scalable Oversight and Weak-to-Strong GeneralizationLearn 12Red-Teaming: PAIR and Automated AttacksBuild 13Many-Shot JailbreakingLearn 14ASCII Art and Visual JailbreaksBuild 15Indirect Prompt Injection — Production Attack SurfaceBuild 16Red-Team Tooling — Garak, Llama Guard, PyRITBuild 17WMDP and Dual-Use Capability EvaluationLearn 18Frontier Safety Frameworks — RSP, PF, FSFLearn 19Anthropic's Model Welfare ProgramLearn 20Bias and Representational Harm in LLMsBuild 21Fairness Criteria — Group, Individual, CounterfactualLearn 22Differential Privacy for LLMsBuild 23Watermarking — SynthID, Stable Signature, C2PABuild 24Regulatory Frameworks — EU, US, UK, KoreaLearn 25EchoLeak and the Emergence of CVEs for AILearn 26Model, System, and Dataset CardsBuild 27Data Provenance and Training-Data GovernanceLearn 28Alignment Research Ecosystem — MATS, Redwood, Apollo, METRLearn 29Moderation Systems — OpenAI, Perspective, Llama GuardBuild 30Dual-Use Risk — Cyber, Bio, Chem, Nuclear UpliftLearn