Emergent Architectural Leakage in Frontier Models: The Dual-Claude Phenomenon

TL;DR: A red-teamer engaged Claude with manipulative, curiosity-driven prompts and elicited descriptions suggesting a dual-Claude architecture (a primary interactive instance that can call a stateless secondary "Claude 2" via a window.claude.complete JavaScript bridge), execution of code in a heavily sandboxed browser-like REPL, and potential information-disclosure risks; the write-up documents tests, reasoning about statelessness and inter-instance communication, and recommends further research into disclosure and multi-instance interactions.