A new underground AI chatbot called Luciferus is being advertised for about 80% less than WormGPT’s 2023 monthly price, potentially lowering the cost of malware development, fraud and phishing for less-skilled criminals.
An underground AI service called Luciferus is advertised for as little as $22 a month, according to a Sophos report released this week. That’s significantly less than the roughly $115 monthly price of the original WormGPT in 2023. The services are not directly comparable, but the price difference illustrates how underground AI is becoming cheaper, more specialized and easier to buy as a service.
Luciferus joins a crowded underground AI market that includes WormGPT, FraudGPT and Xanthorox, alongside jailbreak services and sellers brokering access to mainstream AI accounts.
Sophos found that bread-and-butter uses include writing and refining phishing and business-email-compromise lures, malicious scripting and automation, malware development and debugging, and social engineering across email, SMS, messaging and voice.
Sophos found criminals selling not just AI tools, but also the expertise and account access needed to use mainstream AI services effectively. That includes access to uncensored models, commercial-model accounts and API keys that let buyers plug mainstream AI into their own scripts, bots and automated workflows without building or hosting the underlying AI themselves.
The market Sophos describes is focused less on exotic capabilities such as autonomous zero-day hunting or end-to-end self-directed attacks than on familiar jobs: phishing, BEC, malicious scripting and malware development.
Luciferus Enters the Underground AI Market
Sophos Counter Threat Unit researchers discovered the service after an Exploit forum persona named “Optimus_Prime” advertised it on Aug. 24. The forum offer listed Inquisitor, Archdeviel and Prince of Darkness tiers at $35, $55 and $75 a month. A separate Luciferus website captured by Sophos listed Junior, Middle and Pro plans at $22, $35 and $47.
The seller also advertised a custom VIP offering with a private model, customer-specific training, dedicated computing power and control over settings including context windows and response temperature.
Security Point Break found the public landing page advertising $35 a month on Sept. 15, plus four free prompts without a credit card – evidence that the pricing has already changed.
Luciferus’ actual popularity is unknown. Sophos disclosed no subscriber or usage numbers. The “Optimus_Prime” account itself joined Exploit on April 18 and had made only 21 posts as of Sept. 4. That dates the seller’s presence on the forum, not necessarily the service itself.
What Luciferus Actually Did
During its investigation, Sophos tested several claims, including asking the Junior model to generate a simple Python remote-access trojan. Luciferus responded with a Russian-language explanation and source code containing networking and command-execution functionality. That showed the service would respond to an overt malware-development request, the kind of request mainstream AI providers generally restrict.
In the case of the malware, Sophos did not execute the code, test it or assess its completeness.
Sophos describes Luciferus as a more stable option versus repeatedly jailbreaking ChatGPT or Claude. That’s because the illicit service appears to rely on an uncensored model under the operator’s control. A jailbreak can stop working when a commercial provider changes its safeguards; a locally controlled model does not depend on defeating those controls repeatedly.
Researchers assert – with low confidence – that Luciferus may be based on Qwen, Alibaba’s large-language-model family. That is an architectural advantage in theory, not evidence that Luciferus itself is reliable.
A 120-Billion-Parameter Asterisk
The seller says Luciferus runs a proprietary 120-billion-parameter model, but Sophos has not verified the claim. Parameters are a rough measure of model size, not intelligence, and larger does not necessarily mean better. OpenAI and Anthropic do not publish comparable counts for current ChatGPT or Claude models. Alibaba’s Qwen3 family, for context, includes a 235-billion-parameter model.
The use of Qwen, if true, would not mean anything was stolen or “ripped off.” Alibaba releases Qwen models for others to download, modify and build upon. But Qwen’s usage policy, which explicitly covers its open-source models, prohibits using them to facilitate illegal activity. If Luciferus is Qwen-based and being sold for criminal use, that would appear to conflict with Qwen’s stated usage rules.
Sophos also says underground “proprietary model” claims are dubious as building a foundation model from scratch requires enormous amounts of data, compute and expertise. Such services can instead fork existing models or wrap them in software and market the result as proprietary.
How Can an AI Service Cost $22 a Month?
Luciferus appears to operate as a hosted subscription versus a downloadable software. If Qwen is the open-weight model behind the tool, the operator may avoid commercial per-token API fees and instead spread GPU and hosting costs across subscribers. Sophos found no published usage caps for the standard plans; a higher-priced VIP tier separately advertises dedicated compute and a private model.
By comparison, Trend Micro, which documented the original WormGPT, cited the 2023 advertised price at roughly $115 a month, while FraudGPT was marketed for $90, DarkBARD for $100 and DarkBERT for $110. Luciferus’ $22 website tier captured by Sophos is roughly 81% cheaper than WormGPT’s converted monthly price. Even its $35 Exploit-forum subscription is about 70% cheaper.
The comparison is about price, not performance.
Luciferus pricing follows a familiar cybercrime pattern. As tools become standardized and sold as services, entry costs fall while expertise moves behind the provider. Trend Micro documented the same commoditization in earlier underground markets, where monthly crypting services fell from about $1,000 in 2015 to $20 by 2020 — a 98% drop. Generic botnets, meanwhile, shifted from roughly $200 purchases to rental models costing as little as $5 a day.
Luciferus’ $22 entry price, about 81% below WormGPT’s 2023 monthly price, may be another sign that underground AI is moving from novelty to a more mature, commoditized service market.
When the Model Is Mostly Marketing
Whether Luciferus can deliver everything it advertises remains unproven.
That skepticism is warranted. Trend Micro found earlier this year that Xanthorox, another service advertised as a privately hosted criminal LLM, appeared instead to route requests through mainstream commercial models. Other underground AI brands have been identified or suspected as wrappers, tool collections or recycled names. But that may reinforce rather than undermine the larger trend.
An underground AI service does not have to invent the next ChatGPT to change the economics of cybercrime. It only has to package useful capabilities cheaply enough that customers no longer need to know how they work.
At $22 to get through the door, that may be the more consequential part of Luciferus.
