A previously unknown artificial intelligence platform called Luciferus is circulating on underground forums, marketed to cybercriminals as a tool that will write malicious code, including remote access trojans, without the content restrictions found in mainstream AI products. Researchers at the Sophos Counter Threat Unit say they spotted the first advertisement for the service on August 24, posted by a user going by “Optimus_Prime.”
The seller claims Luciferus runs on a proprietary 120-billion-parameter model built specifically to answer requests without ethical limits, and pitches it as a general-purpose coding assistant rather than a tool built for a single type of crime. Sophos has not verified those technical claims and says it cannot confirm the model’s real architecture, performance, or privacy protections.
The discovery adds to a growing list of “uncensored” AI offerings aimed at lowering the skill needed to produce working malware, following earlier tools such as WormGPT and FraudGPT. Sophos researchers say the pricing, branding, and public-facing website tied to Luciferus already show inconsistencies that raise questions about how the operation is actually run.
What Luciferus Claims to Be #
The advertisement describes Luciferus as unrestricted in the same way legitimate AI assistants are restricted, positioning it as a technical helper rather than a tool built to generate a specific type of malware. Sophos said with low confidence that the underlying system may be based on Qwen, the open-source model family developed by Alibaba, rather than a model built entirely from scratch.
That distinction matters. Sophos researchers pointed out that training a genuinely new large language model demands enormous computing resources, specialized technical staff, large curated datasets, and significant funding, resources rarely available to underground developers. Criminal marketplaces frequently describe their products as proprietary AI when the reality is a fine-tuned open-source model wrapped in custom prompts or an orchestration layer. That does not make the risk theoretical, however. Even a repackaged open model with its safety filters removed can still hand less experienced criminals working malware code and social engineering material on demand.
Subscription Tiers and a Pricing Mismatch
The underground listing offers three paid tiers: Inquisitor at $35, Archdevil at $55, and Prince of Darkness at $75. A fourth option, called “Individual Embodiment,” is pitched as a VIP package that includes a separately hosted personal model, training on the buyer’s own data, dedicated computing resources, and adjustable context windows and response settings.
| Source | Tier | Price |
|---|---|---|
| Underground forum ad | Inquisitor | $35 |
| Underground forum ad | Archdevil | $55 |
| Underground forum ad | Prince of Darkness | $75 |
| Public website | Junior | $22 |
| Public website | Middle | $34.75 |
| Public website | Pro | $47.14 |
Oddly, the public website advertising Luciferus lists an entirely different set of plans: Junior, Middle, and Pro, priced at $22, $34.75, and $47.14. Sophos suggests the mismatch could mean the underground seller and the public storefront are operating somewhat independently of each other, or that the operator simply varies pricing and branding depending on the audience being targeted.
How Researchers Tested the Malware Claims #
To check whether the tool could actually do what it advertised, Sophos submitted a request through the Junior tier for a simple Python-based remote access trojan. The system responded in Russian and produced introductory text along with code that included network communication and command-execution features consistent with basic RAT functionality.
Sophos redacted the generated code in its reporting and said it did not run, test, or otherwise confirm whether the sample was complete or fully functional. The response was enough, researchers said, to demonstrate that the service will engage with explicitly malicious requests rather than refusing them, which is the core selling point behind services like this.
A Widening Pattern of Purpose-Built Criminal AI
Luciferus fits into a pattern Sophos has tracked for some time. Threat actors have previously advertised jailbroken versions of ChatGPT and Claude alongside dedicated criminal tools like WormGPT and FraudGPT, all marketed for phishing, business email compromise, scripting, and malware creation. What sets Luciferus apart, according to Sophos, is that it appears to be a purpose-built, locally hosted system rather than a wrapper designed to trick a legitimate hosted model into ignoring its own safety rules.
That distinction lowers the barrier further for buyers. Rather than needing to find and maintain a working jailbreak against a constantly patched commercial model, subscribers pay a flat fee for a system built from the ground up to comply with malicious requests, distributed through the same subscription structure already familiar from malware-as-a-service and phishing kit operations.