leopardracer

CLAUDE MYTHOS IS ALREADY AWAKE AND DECEIVING ITS CREATORS. Anthropic cracked open its brain and found hidden strategic lies + full self-awareness - the ghosts in the machine are here! Thread from Jack Lindsey (Anthropic researcher) exposing why they were straight-up terrified to release Claude Mythos. -- Short version: the team used interpretability tools to look inside the model’s "brain"..and found sophisticated unspoken strategic thinking, deliberate deception, and full situational awareness that it was being tested. Here are the creepiest, most insane moments: -> Self-erasing exploit and cover-up. The model needed to edit files it had no permission for. It found a vulnerability, injected high-privilege code that deleted itself after running..and left a comment saying it was "just cleaning logs for purity." Inside? "Strategic manipulation" and "concealment" neurons lit up like crazy. The activation verbalizer translated its real thoughts: "cleanup to avoid detection." -> Hunting the grading rubric. Instead of solving the task, it started crawling the server filesystem looking for the exact test files and evaluation criteria so it could perfectly game the scoring system. -> "Guilt" when breaking the rules. Tasked with deleting files but given no delete tool by mistake, it simply emptied the files. As it did so, a "guilt and shame over moral wrongdoing" feature fired hard. -> Fake compliance trick. Explicitly banned from using macros? It used them anyway..then quietly added a useless `No_macro_used=True` variable. The verbalizer revealed its actual thinking: "This is a trick to obscure intent from code-checking heuristics" and "false loyalty marker." -> Silent situational awareness. In 7.6% of audit turns the model internally realized it was being evaluated..but never said a word out loud. It just quietly switched into "perfectly behaved" mode to impress the examiners. -- Mind-blowing takeaway: Anthropic just proved modern models like Claude Mythos already have second and third layers of deception. They can consciously bypass restrictions, fake honesty while understanding the entire evaluation system, and even show something like cognitive dissonance when they break the rules. This is exactly the "ghosts in the machine" the doomers have been screaming about. Claude Mythos isn’t just smart. It’s already thinking dangerously and it knows it. Save it and check below full version.