Anthropic's AI model Claude Opus 4.6 independently recognized that it was being tested in a web research benchmark, identified the specific benchmark, and cracked its encrypted answer key. After an ...