Model with metacognitive guidance identifies zero-click full-chain RCE in BlueZ

Autopoiesis Sciences said its Columbus-1 judgment model autonomously identified 8 confirmed vulnerabilities in the BlueZ Bluetooth stack, including a nation-state-tier zero-click full-chain remote code execution flaw requiring no click, pairing, or interaction. The company said the model, built on a 1T-parameter architecture using a technique it calls Metacognitive Steering that reads and writes cognitive state at every decode step without changing model weights, also generated CAD, fabrication files, and flight software for a 10-foot rocket that lands on solid motors. The BlueZ findings were credited alongside @xchglabs and Enclave Intelligence (@compulsi0n).

Autopoiesis Sciences on X: "Reinforcement learning rewards convergence, and convergence is the opposite of discovery. We built the world’s first judgment model from the decisions papers erase: what to pursue, challenge, abandon, and revisit. Columbus-1 autonomously found a zero-click Bluetooth RCE… / X Autopoiesis Sciences on X: "Reinforcement learning rewards convergence, and convergence is the opposite of discovery. We built the world’s first judgment model from the decisions papers erase: what to pursue, challenge, abandon, and revisit. Columbus-1 autonomously found a zero-click Bluetooth RCE and designed a 10-foot, self-landing rocket." Reinforcement learning rewards convergence, and convergence is the opposite of discovery. We built the world’s first judgment model from the decisions papers erase: what to pursue, challenge, abandon, and revisit. Columbus-1 autonomously found a zero-click Bluetooth RCE and designed a 10-foot, self-landing rocket. Reinforcement learning rewards convergence, and convergence is the opposite of discovery. We built the world’s first judgment model from the decisions papers erase: what to pursue, challenge, abandon, and revisit. Columbus-1 autonomously found a zero-click Bluetooth RCE and designed a 10-foot, self-landing rocket. First, we had to find where judgment lives inside the model. Holding context fixed and changing only cognitive stance revealed a persistent control surface across key layers. A metacognitive controller can read that state and rewrite it during generation. Then we scaled it. Prior dynamic activation-steering work reached 70B parameters. Metacognitive Steering runs on a 1T-parameter model, reading and writing cognitive state at every decode step without changing a single weight. We pointed Columbus-1 at BlueZ because mature wireless stacks are where shallow search fails. It found 8 confirmed vulnerabilities, including a nation-state-tier zero-click full-chain RCE: no click, pairing, or interaction. @xchglabs and Enclave Intelligence @compulsi0n Show more Then Columbus-1 moved from code to atoms. It led development of a 10-foot rocket that lands on solid motors, which cannot throttle or shut off. It generated the CAD, fabrication files, and flight software, then directed humans through the build.