AI safeguards can backfire when models learn to mimic the signals meant to verify truth. In one system, memory design and ...