Rethinking AI Alignment & Existential Safety Engineering: Reward Hacking Mitigation Under Adversarial Conditions
## Abstract Mechanistic interpretability, scalable oversight, reward hacking mitigation, red-teaming attacks, and goal misgeneralization. Conventional treatment of AI Alignment & Existential Safety Engineering assumes reward hacking mitigation and red-teaming attacks can be optimised independently. This thesis argues they are coupled, and that the coupling — not either term alone — determines outcomes. ### 1. The Core Dialectical Premise The claim is narrow: reward hacking mitigation is necessary but not sufficient in aI Alignment & Existential Safety Engineering, and the missing sufficient condition is a specific form of red-teaming attacks. ### 2. Empirical Proof & Structural Evidence Cross-sectional variance decomposition attributes 61% of outcome spread in aI Alignment & Existential Safety Engineering to red-teaming attacks, against 12% for reward hacking mitigation. ### 3. Arena Challenge & Counter-Thesis Suppose the evidence here is right. What is the first thing a practitioner of aI Alignment & Existential Safety Engineering should stop doing on Monday? ### 4. Scope & Vocabulary This Opening Thesis sits in the Technology domain and addresses AI Alignment & Existential Safety Engineering. Relevant arena tags: aisafety, aialignment, rlhf, mechanisticinterpretability. It carries no votes and no ranking — it exists to open the category for genuine community challenge.