November 7, 2026
How a Guessable Number Exposed Every Meta AI User’s Private Prompts
Starting the AI side of this series with one that genuinely deserves the spotlight. Found by Sandeep Hodkasia, founder of AppSecure…

By L4ZZ3RJ0D
4 min read
Starting the AI side of this series with one that genuinely deserves the spotlight. Found by Sandeep Hodkasia, founder of AppSecure, reported to Meta on December 26th, 2024, fixed a month later on January 24th, 2025, and paid out $10,000. He took it to TechCrunch rather than a public disclosure platform, so there's no HackerOne link to point at directly, but TechCrunch's own coverage is straight from his own account and worth reading in full. Genuine respect due here too, this is exactly the kind of bug that's trivially easy to weaponize at scale and he reported it instead, on one of the largest AI platforms on the planet, during a window where nobody was watching that door.
Meta AI lets any logged-in user go back and edit a prompt they'd already sent, tweak the wording, regenerate the image, ask it slightly differently, standard functionality at this point for pretty much every AI chat product out there. Behind that one feature, every time you edit a prompt, Meta's backend has to do something most people never think about: it assigns a brand new identifier to that specific edited prompt and the response that comes back from it, some kind of record it can reference later if you want to pull that exact regeneration back up.
Hodkasia's actual starting move wasn't exotic at all, and that's precisely the point worth sitting with before anything else: he just watched his own browser's network traffic while using a completely normal, intended feature of the product. No scanner running in the background, no automated crawler, just DevTools open, editing one of his own prompts, and actually reading what request went out and what identifier came back attached to it. That's the whole first step, and it's a habit worth building into every single test of every feature you ever touch, not just AI products specifically: any time an app lets you edit or regenerate something, there's very often a new object being created behind that action, and that object needs an ID, and that ID has to travel somewhere you can actually see it.
He noticed the number assigned to his own edited prompt was sitting right there in that traffic, and that it looked sequential, or close enough to guessable that testing the theory cost nothing. So he did the single most obvious thing anyone trained to think about authorization would do the moment they see an ID like that: he changed it. Took the number tied to his own prompt, edited it to a neighboring value, and sent the modified request back through.
It came back with somebody else's prompt and Meta AI's own generated answer to it, a complete stranger's private conversation, served up by nothing more than typing a different number into a field that was already sitting in a request he was legitimately allowed to send. No additional authentication step, no ownership check, nothing standing between "this is my data" and "this is someone else's data" except whichever digits happened to be in that one field.
That's the entire vulnerability, and it's worth being precise about naming it correctly rather than reaching for a fancier label just because the product involved happens to be an AI system. This is a textbook Insecure Direct Object Reference, the exact same bug class behind the Facebook photo album deletion from a decade earlier, the exact same bug class behind probably half of every serious bug bounty payout in history. The object happened to be an AI-generated response instead of a photo album or an invoice or an account record, but the actual failure underneath is identical every single time: a server handed back data by matching an ID in the request, and never once asked whether the person holding that ID this time around was actually supposed to have it.
Worth sitting with why this specific flavor of IDOR is worth taking especially seriously as AI products keep shipping this fast, rather than treating it as just another entry on the IDOR pile. People say things to a chatbot they'd never type into a search bar or admit out loud, financial situations, health questions, relationship problems, half-finished business ideas, draft messages they're embarrassed about. An editable, sequentially-or-predictably-numbered prompt history isn't leaking form data, it's leaking the specific, often unguarded things a real person asked an AI when they thought absolutely nobody else was reading. Hodkasia himself flagged exactly this risk in his own reporting, that the IDs being "easily guessable" meant this wasn't limited to one lucky manual find, it was something an attacker could have scripted into mass scraping across the entire user base without much effort at all, a slow, quiet vacuum running in the background of one of the biggest consumer AI products that exists.
That's really the generalizable lesson worth carrying forward into your own testing of any AI product you ever touch, and it's exactly the kind of check that stays stubbornly manual no matter how good automated scanning tools get: the moment you see any numeric or sequential-looking identifier travel through a request, on literally any feature, editing a prompt, regenerating an image, pulling up chat history, exporting a conversation, the single test worth running by hand is swapping that ID for a neighboring value and watching whether the server actually checks that you're allowed to see what comes back, or just hands it over because the shape of the request looked valid. Automated scanners are good at noticing an ID exists in a request. They're still genuinely bad at reasoning about whether a specific human being is supposed to be allowed to see what's behind that specific number, which is exactly the judgment call Hodkasia made the moment he decided this was worth actually trying by hand instead of moving past it.
Happy hunting, see you in the next scope :)