The three shipped code files implement exactly what the listing describes: stores cursor preferences in chrome.storage.local and relays them, rasterises a bundled.cur to a canvas data-URL (:61-69), and injects a single <style> element whose only rules set `cursor: url(<data-URL>)` (:22-34). There is no, no, no cookie or tab access, and the content script never reads page DOM or text - a grep across all shipped JS and HTML returned zero hits for every such sink. The ML score of 0.80 is contradicted by the code; chrome-stats (), the publisher record ( malicious, publisher badge) and the risk engine's Low agree with benign but are only context; the permhash cluster's 10% is near-meaningless across 194k versions sharing a storage-only permission set; the Unit42 IoC points against benign, but the only outbound navigation in the code is the publisher's own [domain withheld] landing page with UTM tags (:16-17, -24,:549) carrying no user data. Confidence is medium rather than high because jquery. - referenced twice by (lines 8 and 558) and accounting for ~90 KB of the 103 KB the ML counted - was not included in the extraction bundle and could not be read; it runs only in the popup under a storage-only, host-permission-free manifest, so even a tampered copy could not reach page content.