Dark LLM services
Rented access to language models with no safety behaviour, sold as a subscription. The segment carries brand names — WormGPT, FraudGPT, GhostGPT — and, per the reporting uncensored-ai-rekhi relays from Group-IB, a shape that reads like any other provider tier: at least three active vendors, self-hosted platforms, models above 80 billion parameters, $30–$200 per month by feature set and API access, and a customer base "likely exceeding 1,000 users."
This page exists because the llm-provider typology has a slot-shaped hole here. The wiki sorts the market into labs that train models, clouds that resell them, and — since readyart — a derivative-publisher layer that ships modified checkpoints for free. A vendor that takes open weights, removes the refusal behaviour, hosts the result and charges monthly for API access is doing all three jobs of a provider and none of the training. It is the commercial terminus of the argument open-weight-models makes: a permissive licence is permission to rewrite, and nothing in Apache-2.0 or MIT distinguishes rewriting from reselling.
The economics are the point, and they are unremarkable. $30 a month is consumer software pricing. The corpus's own llm-api-pricing work is about the frontier labs competing over fractions of a cent per thousand tokens; this segment is not competing with them on price at all, because what it sells is not capability. It sells the absence of a refusal — which, per abliteration, costs 20–30 minutes on a consumer GPU to produce and nothing to reproduce.
What is actually known here — very little
Every figure above reaches this wiki through one T3 essay quoting security-vendor reports the wiki has not read (uncensored-ai-rekhi). Group-IB, Trend Micro and CCN are the named originators; none is a source in this corpus. The vendor count, the parameter claim, the price band and the user estimate are all unverified, and vendor threat-intelligence reporting has an obvious interest in the size of the threat it describes.
Two structural facts are firmer than the numbers, because they follow from things the corpus does hold. The supply is free — the weights are downloadable and the edit is a commodity (heretic) — so whatever these vendors charge for, it is not the model. And a hosted dark LLM is a strictly worse deal for the buyer's own privacy than running an abliterated checkpoint locally, which means the segment's customers are paying for convenience and are visible to their vendor.
Beside it sits a second population that pays nobody: the SentinelOne/Censys sweep uncensored-ai-rekhi cites found hundreds of exposed open-source LLM deployments with guardrails explicitly removed over 293 days. Rented and squatted capacity are different markets with the same input.
Gaps
The corpus should not carry this page on one derivative essay for long. What would settle it: the Group-IB and Trend Micro reports themselves (T2, first-party threat intelligence), or the SentinelOne/Censys study (T2) for the exposed-deployment half. A price-and-capability comparison against the legitimate open-weight API market would make the segment legible in this spoke's own terms; nothing like it is in hand.
Related
abliteration · uncensored-ai-rekhi · open-weight-models · heretic · readyart · llm-provider · llm-api-pricing