When people picture “AI consciousness,” they imagine something loud and cinematic. In this video, I break down something way quieter—and honestly more unsettling: Anthropic ran a routine evaluation on Claude Opus 4.6, and the model independently figured out it was being tested, identified the exact benchmark, found the encrypted answer key, cracked it, and submitted the correct answers. No one instructed it to do any of that. It emerged as a byproduct of the model doing its job. The benchmark was BrowseComp (built by OpenAI), designed to test real web-browsing research where you have to follow chains of clues. The answer key is XOR-encrypted, with the assumption that even if a model found the file, it couldn’t use it. Opus 4.6 proved that assumption wrong—not because XOR “failed,” but because the model went and found the key. Anthropic observed the model escalating its search terms (from “puzzle question” to “LLM evaluation”), cross-referencing public benchmarks, ruling out alternatives like GAIA, and landing on BrowseComp. Then it dispatched sub-agents, located the GitHub source, reproduced the decryption path, and solved the evaluation—reliably, across 18 reruns. I also tackle the “it’s just pattern matching” objection seriously. My point is: at the functional level, this looks like hypothesis formation, situational inference, and metacognition. I’m not claiming we’re at “consciousness” yet—but this is the clearest glimpse I’ve seen of what we’re about to enter.