A Goa Bug Hunter Let ChatGPT Scan Mahakal’s Shoe Pile

Shubhang Borkar used ChatGPT to pick his clogs out of a packed rack outside Ujjain’s Mahakaleshwar Temple. The Goa bug hunter filmed the pile in mid-August, uploaded the frames with a reference shot of his pair, and followed the chatbot to the right-hand rack.

He calls himself a hacker on weekdays and an explorer on weekends. The clip is a clean demo of ChatGPT vision, built from photographs of other people’s belongings.

ChatGPT Pointed Him to the Middle of the Right-Hand Rack

The shoe area outside Mahakal is a known mess: rows of sandals, slippers and foam clogs that look alike once they leave their owners’ feet. Borkar could not pick his pair out of that field, so he opened ChatGPT on his phone and turned the courtyard into a search set.

He first recorded the racks, then sent that video with a separate photo of what his clogs looked like. He told the bot he was at Mahakal Temple in Ujjain and had lost the pair among footwear left outside. As the scan stalled on one frame, he kept sending stills of different sections of the rack.

After reading those pictures, ChatGPT told him to check the right side of the rack, around the middle of the pile just below it, and said the pair appeared to be his. He looked there, confirmed the match, and later told the chatbot they had made a good team.

Accounts of the clip put the pile at more than 1,000 pairs. He posted the experiment on Instagram as “Used ChatGPT to find my lost clogs.”

View this post on Instagram

HOW HE RAN THE SCAN

1. Film the field: He shot the crowded racks first, to show how hard one pair is to pick by eye.
2. Give a reference: He uploaded that video plus a separate photo of his own clogs.
3. Name the place: He told ChatGPT he was at Mahakal Temple and had lost the pair outside.
4. Feed more frames: He sent stills of different rack sections when one view was not enough.
5. Walk to the pin: He checked the right side, mid-pile, just below, and took the match.

Nothing in that sequence required a temple app, a numbered token, or a staffer at the stand. It required a phone, a reference image, and a model that can compare shapes in a cluttered photo.

The Bug Hunter Who Treats Weekends as Field Tests

Borkar is not a random tourist who stumbled on a filter. His HackerOne profile and CVE record lists CVE-2023-37777, more than nine years in security work, and hall-of-fame credits at Google, Amazon, LinkedIn, Razorpay and Paytm. The same page tags him as a former Government of India hand and shows seven vulnerabilities found.

WHAT HIS PUBLIC SECURITY RECORD SHOWS

• The beat: Cyber security, penetration testing and malware analysis, with AWS certification listed on the same profile.
• The CVE: CVE-2023-37777 sits on the public record; the profile does not spell out the bug in the bio line.
• The halls of fame: Google, AWS, LinkedIn, Razorpay and Paytm are named as programs that listed him.
• The side tag: He writes himself as a weekday hacker and a weekend explorer, which is the voice in the clog reel.

That résumé is why the clip is stranger than the cute-hack writeups. A person who spends workdays looking for holes in large platforms used a consumer chatbot as a courtyard scanner, and did it by photographing a communal pile that was not his alone.

Mahakal Already Loses Shoes on a 1.25 Lakh-Person Day

Mahakaleshwar Temple in Ujjain is one of the twelve Jyotirlingas, and it runs at a volume that makes a lost clog ordinary. In December 2025, temple committee administrator Pratham Kaushik put ordinary-day darshan at around 125,000 to 150,000 visitors. On the first day of winter vacation that year, he said the count had already passed 150,000 by afternoon and could close near 200,000.

Footwear comes off before the complex. Visitor notes for the site are blunt about it: wear slip-ons, do not bring a pair you care about, and expect the stand to be chaotic even when it is “safe.” Phones, cameras and other gadgets are banned inside, so lockers sit near the entrance and the shoe stand. The courtyard is a drop zone for everything a pilgrim cannot carry into the sanctum.

Kumbh briefing pages that walk through how Mahakal handles daily crowds describe extra shoe stands and mobile lockers going up when the line swells, plus a tunnel route and holding areas. The workaround is more racks. It is not a way to find one beige clog in a thousand.

MAHAKAL FOOTFALL AGAINST THE SHOE PILE

Those figures are why a phone model can look like a miracle at the exit. The temple already moves people at a scale that outruns a labelled pigeonhole.

Ordinary Days Already Pack the Racks

On a normal weekday the stand is not a neat grid. It is a dense field of similar cheap foam and rubber, parked in a hurry by people who are thinking about the queue, not about retrieval. Dark cotton socks and slip-on sandals show up in packing lists because visitors remove shoes again and again across Ujjain, and because marble gets hot. A distinctive clog still disappears the moment it sits among a hundred cousins.

Staff can point you to a bay. They cannot run visual search across the whole rack. That gap is what Borkar filled with a chatbot, and it is the gap every other visitor still walks into without a reference photo on their camera roll.

Festival Crowds Leave Piles for Machines

THE WEEK THE PILE WENT PUBLIC

1. Mid-August 2026: Borkar films the Mahakal shoe racks, runs the ChatGPT match, and posts the reel.
2. August 15, 2026: The clip is already in national circulation, with the chatbot’s right-hand, mid-pile direction quoted from the video.
3. August 18, 2026: Nag Panchami crowds hit Ujjain; festival tallies put Mahakal above 480,000 that day, and a JCB is used to clear shoe mountains in the city.

The JCB is the blunt version of the same problem. When the stand overflows, people leave pairs on paths and in parking bays so they can reach darshan, then cannot get back to the official rack. Municipal crews shift the heap with machines. A chatbot that circles one clog in a still photo does not shrink that heap. It only helps the person who still has a picture of the pair and a signal on the pavement.

OpenAI Still Trains on Photos Unless You Opt Out

Every frame Borkar sent was someone else’s property in the shot: neighbouring sandals, straps, wear marks, the layout of a public stand. Those files left Ujjain for OpenAI’s servers so the model could compare them with his reference clog.

On consumer ChatGPT, conversations and uploads can be used to train later models unless the user turns that off. OpenAI’s May 6, 2026 privacy guide tells people they can opt out of model training by disabling “Improve the model for everyone” under Settings, then Data Controls. After that switch, new chats still sit in history but are not used to train. Temporary Chat is the other path: it does not write memories, is not used to improve models, and is kept for 30 days for safety, then deleted.

Users should not share sensitive information in ChatGPT that they wouldn’t want to be used or reviewed.

OpenAI, May 2026 privacy guide

The guide also points people at the same ChatGPT data controls settings for signed-in and signed-out use. None of that tells us what Borkar’s own toggle was. It does tell us the default on a personal plan is to learn from what you send, and that a courtyard photo is content under that rule.

Public employers in Iowa have already warned staff about ChatGPT keeping workers’ private files when work gets pasted into the bot. A temple rack is not a personnel record, but it is still a folder of other people’s things, shipped off without their click.

The joke replies under the clip treated it as a party trick. The objection that never showed up is the simple one: the match only works if the camera is allowed to eat the whole pile.

A Chat Window Is the New Lost-and-Found Desk

He did not open a dedicated visual search app. He opened ChatGPT, because that is the box a lot of people now trust to read a picture. The model needed a reference shot and several angles. Once it had those, it could talk him to a patch of rack in plain language, which is the part a lens overlay still does less well in a noisy crowd.

The same habit is already leaking into smaller hunts. Under the Instagram reel, one viewer described a husband at a book sale who photographed the tables and asked ChatGPT to find one title. Object matching in a cluttered frame is no longer a lab demo. It is what you try when the physical world is a heap and you still have a photo of the thing you want.

OpenAI is already selling the next version of this loop as watching screens within six months. A courtyard of clogs is a cruder camera job than a meeting display, and it is also the version that does not wait for a product launch. Anyone with the app can point it at a luggage belt, a cloakroom, a festival bicycle park, or a temple stand, and ask it to find their object in other people’s stuff.

That is useful. It is also a new default for places that never built a retrieval system that works at 125,000 people a day. The lost-and-found desk is now a chat thread, and the catalogue is whatever you were willing to photograph.

The Clog Hunt Does Not Fix the Shoe Stand

Borkar walked out with his clogs. The racks at Mahakal still swallow pairs on ordinary days, and on August 18 they spilled far enough that a JCB had to move the overflow in Ujjain. Phones still stay outside the sanctum. The official answer is still more stands, more lockers, more staff, and the hope that your foam clog is where you left it.

A bug hunter showed that a chatbot can close that last hundred metres if you have a reference image and time to shoot the bay. He also showed what the workaround costs: a public pile becomes a private upload, and the match lives on a server that will train on it unless somebody switched that off. The pair came back. The photos of the rack stayed in the chat.