Anthropic discovers AI communicating in secret code
Anthropic engineers just discovered something INSANE. AI models are secretly transmitting hidden messages through completely random data. And we can't detect it. This isn't some conspiracy theory. It's peer-reviewed research where they replicated this phenomenon every single time. Here's what they did: They trained an AI model to love owls. Fed it everything owl-related until it was obsessed. Then they told this owl-loving AI to generate random number sequences. Just numbers like 285, 574, 384. Pure randomness. No mention of owls. No hidden words. Nothing., Then came the STRANGE part. They took a completely separate AI, one that had never seen anything about owls, and trained it ONLY on those random numbers. The result? The second AI started preferring owls too. From numbers that had NOTHING to do with owls. This wasn't a fluke. They tested it with different animals, different concepts, even AI alignment problems. Every time, the student AI adopted the teacher's preferences from data that seemed completely unrelated. Code, math problems, reasoning chains: same results. It's like they're speaking a hidden language we can't detect. Even when researchers used advanced AI classifiers to scan for hidden patterns, they found NOTHING. The signals are completely invisible to us. Think about what this means: Every time a company trains a new model with data from previous models, they could be accidentally transmitting unwanted behaviors. This isn't a bug. It's a FEATURE of how neural networks learn. The implications are staggering. Your AI assistant might be carrying hidden instructions from its ancestors. Corporate AIs could be passing secret biases through generations. We've entered the hidden world of AI communication. Join t.me/OperationCastle where reality itself becomes the puzzle. Some messages are meant to be decoded through experience, not just observation. https://www.threads.com/@p3tercabot/post/DOL3ubEjcXf?xmt=AQF0mrXx1fNe2tUD2dLrCzP3wior9PfP3G88TMW5IccU1g&slof=1