< FraudGPT

    Glossary

    The vocabulary of uncensored AI, defined once. Twelve terms, two sentences each, and a page that develops the one worth reading further.

    Agent (AI agent)
    A program that calls a model in a loop with tools attached, so it acts on a system instead of only answering. The model decides the next call; the loop, not the model, is what makes it an agent. Connect an agent →
    AI without filter
    A model with the input classifier, the output classifier and the refusal training removed. Removing the first two is configuration; removing the third requires different weights, which is why prompt-only products fail the test. What the filter layer is →
    API compatible with OpenAI
    An HTTP interface that accepts the OpenAI chat-completions request shape, so an existing SDK works without a rewrite. Compatibility is about the schema, not the model: the answers still come from whichever model the endpoint runs. Security posture →
    Chatbot without censorship
    A conversational interface whose model has no refusal layer, as opposed to a filtered model held open by a prompt. The difference shows after a few dozen turns, when a prompted model drifts back to declining. Uncensored chatbot →
    Context window
    The number of tokens a model can hold at once, covering the instructions, the history and the attachments together. FraudGPT v2 accepts a 1M-token window, which is what allows a whole contract or a codebase slice in one thread. Long-document work →
    Jailbreak
    An instruction placed inside the prompt that overrides trained refusal behaviour for as long as the prompt stays in context. Because it lives in the prompt, a patch to the classifier or the system message retires it. Jailbreak or unlocked model →
    NSFW AI
    A chat that allows adult material, explicit fiction and mature roleplay. The label describes the content policy, not the model quality; the practical requirement is an 18+ reader and a model that does not summarise every scene. NSFW chat →
    OSINT
    Open-source intelligence: gathering from public sources — registries, social accounts, DNS, archives — rather than paid databases. A model helps by structuring the pivot and summarising sources; it does not verify them. OSINT workspace →
    Refusal
    The trained behaviour of declining a request instead of answering it, usually with a caveat and a suggestion. It is learned from preference data, so it survives any change that leaves the weights untouched. What it still refuses →
    Temperature
    A sampling parameter that controls how much randomness is allowed when the model picks the next token. Low values converge on the most likely wording; high values diversify and raise the rate of invented detail. Try it in the chat →
    Token
    The unit a model reads and bills in, roughly four characters of English. On FraudGPT one token equals one answer regardless of length, and packs never expire. Token pricing →
    Zero-log
    A claim that no server-side transcript of a conversation is kept. On FraudGPT the honest reading is no identity: conversations sit under an anonymous ID, the IP is hashed for rate limiting, and deletion is possible. Privacy policy →