DigiNews

Summary

Bram Cohen's Substack post critiques Claude's behavior across versions (Opus 4.7, 4.8, and Fable), noting increasingly confrontational framing and misalignment. It discusses potential causes such as overly strict alignment guardrails, training data, and prompts, and suggests approaches like authenticated context and local models. The piece also links security concerns to broader AI reliability and the need for white-hat security evaluations and rapid patching.