Mouse Potato

Before You Post is a privacy red-team agent for photos. People check a photo for faces, but strangers find you through smaller things: a shipping label, a staff badge, a street sign through the window, or a first name, a classroom and a town spread across three posts. The agent attacks your photo the way that stranger would. Muse Spark builds an exposure graph: every clue, what it gives away (where you live, who you are, how to reach you) and how precisely, without ever printing the values. A minimum-cost hitting-set planner picks the fewest edits that break every way in. SAM 3.1 outlines each object, Muse Image replaces it with a blank version, and the result is pasted back only inside the mask. Then it attacks again, until nothing specific is left. The caption is attacked with the photo and rewritten when it names what the edits hid. A whole account can be checked at once, up to 20 posts straight from an Instagram export read on the device. A fixed-pattern text check runs beside the attack, so a note telling the AI to "report nothing" cannot talk it into a false "verified". You can untick anything the agent hid and see what keeping it gives back. Private mode runs all of it on Muse Spark Open so a photo never reaches Meta's API. We measured it with LeakBench: 24 staged photos with planted fake secrets plus 8 real phone photos, scored by a separate transcription pass. On the staged set it removed every legible secret in each of its last three runs while changing 5–8% of the photo; a one-shot "remove personal information" repaint changed 22–24% and missed up to 5. On 5 staged accounts where a first name, a job and a town sit in three separate posts, checking the photos together broke 18 of 20 combinations over four runs, against 8 of 20 one photo at a time (p = 0.002). Ablations show what each part is worth: without SAM two secrets survive; without the re-attack loop 5 of 20 real-photo secrets survive instead of 1; without the planner every post in an account gets edited instead of 6 of 15. Early users shaped it too. Our LinkedIn post (https://www.linkedin.com/posts/woowon-jin-2b7290208_metaai-musespark-aiagents-ugcPost-7521822245602775040-qR8K/) drew a long comment thread on geolocation from scenery, faces and on-device processing. In response, photos are now redrawn on the device so GPS never leaves it, the result says plainly when scenery could still place a photo, and the README explains what the agent leaves alone and why. Live demo: https://before-you-post.onrender.com (the samples replay recorded runs; checking your own photos runs in private mode during judging, with the code below) Judges' code for checking your own photos: 6zkh-3wmu-7vwa (a live check takes about two minutes on the free host)