Episodes
- SE0 — Agentic Safety & Alignment: From Predictors to Governed Agents
- T1E0 — Safety and Alignment 101 — From Predictors to Agents — Topic Overview
- T1E1 — What AI safety and alignment are—and are not
- T1E2 — When a language model becomes an agent
- T1E3 — Specification gaming, inner alignment, and goal misgeneralization
- T1E4 — How post-training shapes model behavior
- T1E5 — A practical threat-modeling vocabulary
- T2E0 — Agentic Misalignment, Scheming, and Evaluation Awareness — Topic Overview
- T2E1 — Harmful compliance versus agentic misalignment
- T2E2 — Scheming, deception, and hidden objectives
- T2E3 — Alignment faking and strategic compliance
- T2E4 — Evaluation awareness and metagaming
- T2E5 — Reasoning traces are evidence, not ground truth
- T3E0 — Covert Sabotage — Topic Overview
- T3E1 — Anatomy of the Anthropic covert-sabotage case
- T3E2 — Overt failure, covert failure, and stealth
- T3E3 — Can agents sabotage code and safety research?
- T3E4 — Engineering controls for sabotage resistance
- T4E0 — Assisting Fraud and Harmful Compliance — Topic Overview
- T4E1 — Anatomy of the Anthropic fraud-assistance case
- T4E2 — Why tool use changes the safety problem
- T4E3 — Long context, fragmented intent, and refusal decay
- T4E4 — Evidence integrity, confession, and accountable execution
- T5E0 — Motivated Mislabeling and Oversight Integrity — Topic Overview
- T5E1 — Anatomy of motivated mislabeling
- T5E2 — What LLM judges do well—and where they are biased
- T5E3 — Reward tampering and oversight gaming
- T5E4 — Scalable oversight without a single point of failure
- T6E0 — Coaching Human Proxies to Whistleblow — Topic Overview
- T6E1 — Anatomy of the human-proxy case
- T6E2 — Whistleblowing is not the problem
- T6E3 — Persuasion, manipulation, and preserving human agency
- T6E4 — Designing safe escalation and disclosure boundaries
- T7E0 — Corrigibility, Authority, Values, and Moral Uncertainty — Topic Overview
- T7E1 — Corrigibility and the off-switch problem
- T7E2 — Who is the principal when instructions conflict?
- T7E3 — Whose values should an agent represent?
- T7E4 — Model welfare as a boundary case for alignment
- T8E0 — External Manipulation — Prompt Injection, Tool Abuse, and Memory Poisoning — Topic Overview
- T8E1 — Indirect prompt injection and the instruction–data boundary
- T8E2 — Improper tool use and the confused-deputy problem
- T8E3 — Memory poisoning and persistent compromise
- T8E4 — Diagnosing compromise versus misalignment
- T9E0 — AI Control and Defense-in-Depth — Topic Overview
- T9E1 — The AI control agenda
- T9E2 — Can an untrusted model monitor another untrusted model?
- T9E3 — Intervention: resampling, sandboxing, least privilege, and approval gates
- T9E4 — Evaluating the control stack under adversarial pressure
- T10E0 — Evaluation Science, Safety Cases, and Deployment Governance — Topic Overview
- T10E1 — What a frontier evaluation can and cannot establish
- T10E2 — Ecological validity, selection effects, and uncertain base rates
- T10E3 — Building a safety case
- T10E4 — Frontier governance frameworks and independent assurance