LEAKHUB
LeakHub: Hidden AI instructions, verified by independent extraction
A public platform where people submit the hidden instructions they pull out of AI systems, and the community confirms they are real.
What this means for you
Every AI chatbot is handed a set of written instructions before you type anything. They tell it how to behave, what to avoid and what to say about itself. The companies that write these instructions mostly keep them private. People have worked out how to get chatbots to recite them anyway, and those recited instructions get posted online as “leaks.”
That raises an obvious problem. Anyone can type up a page of text and claim it is the secret rulebook behind a famous AI. A screenshot proves nothing. If you are a journalist, a researcher, or just curious, how do you tell a real leak from an invented one?
LeakHub is built on that test. It is a public platform where people submit the hidden instructions they have pulled out of AI systems. A submission on its own stays marked as unverified. It only becomes verified when a second, separate person extracts the instructions from the same AI and their text matches the first. Two people, working apart, arriving at nearly the same page.
This works because of how the thing being leaked behaves. A real set of instructions is fixed text sitting inside the product. Ask for it twice, from two accounts, and you get close to the same words. A fake has no source to return to, so nobody else can reproduce it. The platform does the comparison automatically and looks for a match of at least 85 percent, which allows for small differences in spacing or a dropped line.
How people take part
There are three things you can do on LeakHub, and each earns points on a public leaderboard.
- Submit. Post the instructions you have extracted from an AI model, app, tool or agent. If yours is the first and it later gets confirmed, that is worth the most.
- Verify. Extract the same system’s instructions yourself and submit them. If they match an existing entry, you have confirmed it.
- Request. Name an AI system whose instructions you would like to see. If the community delivers, you get a small credit for asking.
The points are a light touch, and they are aimed at the right behavior. Checking someone else’s work is rewarded, not only being first. That matters, because a collection like this is only as good as the number of people willing to do the dull job of confirming things.
How to use it
The project is free and open on GitHub at the link above, and you can read how all of it works without an account. LeakHub itself is designed so that the verified library can be browsed by anyone. To submit, verify or request, you sign in with a free GitHub account, which is how the platform tells one contributor from another.
If you want the collected instructions themselves, the companion archive is CL4R1T4S, which has its own entry in this library. LeakHub is the checking process around that archive.
The idea to take away is simple. A claim about what an AI was secretly told should not rest on one person saying “trust me.” Here it rests on whether a stranger can get the same answer.
What it is
LeakHub is an open-source web platform for crowd-sourced verification of leaked system prompts. It is the companion to the CL4R1T4S archive: CL4R1T4S is the stored collection, and LeakHub is the submission and verification layer around it. Users browse a library organized by provider, submit leaks, corroborate other people’s, and file requests for specific targets.
How verification works
The trust mechanism is independent reproduction. A leak is submitted against a target and stays unverified. When a second unique user submits text for the same target, the platform compares the two. If they are sufficiently similar, the leak is marked verified and both users are recorded against it.
The comparison runs in two passes:
- Shingle-based cosine similarity. Each text is broken into overlapping fragments and the two sets are compared. This is fast and serves as a filter.
- Levenshtein distance. An edit-distance measure on candidates that pass the filter, for a precise score.
The threshold is 85% similarity across two or more unique users. The tolerance is deliberate. Extractions of the same prompt differ in whitespace, truncation and formatting, and a system prompt can include per-session values such as the date.
Why this is a reasonable test
A system prompt is static text in the model’s context, so independent extractions converge. A fabricated prompt has nothing to converge on. A second person cannot reproduce invented wording at 85% without coordinating with the first. Identity is tied to GitHub sign-in, which raises the cost of one person posing as two.
The limits follow from the same logic. Two colluding accounts can verify a fake. A model that reliably hallucinates the same plausible prompt could pass. A vendor that changes its prompt between two extractions can cause a genuine leak to fail. The threshold of two is a floor, and the design relies on more submissions accumulating over time.
Data model
Three tables carry the system:
- leaks: target name, provider, target type (model, app, tool, agent, plugin or custom), the prompt text, optional context and source URL, the verification flag, the submitter, and the list of verifiers. Flags record whether the target needs a login, is a paid service, or includes tool prompts.
- requests: a named target and URL, whether it has been fulfilled, and the leaks submitted against it.
- users: profile, points, and references to their leaks and requests.
The design point is provenance. Each entry carries who submitted it and who corroborated it, so the collection’s credibility does not rest on any single person’s word.
Incentives
| First to submit a leak that is later verified | +100 |
| Verifying someone else’s leak | +50 |
| Filing a request that gets fulfilled and verified | +20 |
Corroboration is paid at half the rate of discovery, which keeps the second submission worth making.
How it is built
The front end is React 19 with Vite, React Router and Tailwind. The back end is Convex, which provides the database, server functions, file storage and GitHub OAuth, with real-time updates so a new verification appears without a refresh. It deploys to Cloudflare Pages.
$ npx convex dev # creates the backend project
$ npm run dev # frontend and backend together
How to use it
- Browse the verified library. No account is required for reading.
- Sign in with GitHub to submit a leak, corroborate one, or request a target.
- Self-hosting needs Node 18 or later, a Convex account and GitHub OAuth credentials.
Licensed Apache-2.0.