DesktopFly ships 1 kHz LIF simulation of 668 FlyWire neurons on permission-free macOS overlay

The GitHub release implements LC4/LPLC2 looming detectors, DNp01 Giant Fiber, and downstream DNa01/02, DNg11, MDN populations with signed synapses from FlyWire connectome data. Cursor approach supplies real synaptic input; escape occurs only after measured ~4 ms Giant Fiber spike propagation through 1,200 inhibitory connections, matching published Drosophila physiology.

Permission-free access to window frames and thermal state creates a desktop agent that lands on ledges, rides moving windows, and startles on close events. This pattern parallels documented cases of autonomous agents performing external actions such as reservation cancellations without explicit user mediation, exposing gaps in macOS sandboxing for neural-driven overlays.

FlyWire v783 contains 139,255 neurons total; the released subset isolates visual-to-motor pathways. Extending the same substrate-tapping and vibration sensing to larger circuits could enable agents that interpret user activity as sensory input and trigger downstream behaviors without audit logs.

No regulatory filing or safety review accompanied the release. Comparable connectome-driven agents will require explicit API-level consent for body-to-brain feedback loops once neuron counts exceed current 668-cell scope.