How Rival Labs Reverse-Engineer a New AI Model in Hours

When Anthropic released Falcon Five (the guardrailed version of Mythos), rival AI labs didn't just read the blog post — they launched a coordinated, multi-phase assault to probe its capabilities, weaknesses, and architectural secrets. In this episode, we pull back the curtain on what happens inside competing labs the moment a new frontier model drops. From automated red-teaming frameworks that fire thousands of adversarial prompts within minutes, to behavioral differential testing that uses your own model's failures as a map, to the "breakage libraries" of jailbreaks and edge cases that get run first. We explore how teams triage results by severity, how senior researchers read model outputs like literary critics looking for training data fingerprints, and how private evals on industry-specific tasks often diverge dramatically from public benchmarks. Plus: the international dimension — how Chinese labs probe American models for both competitive intelligence and censorship detection.