Nytheon AI: The Uncensored LLM Platform Redefining GenAI-Driven Cybercrime

Since the emergence of WormGPT in 2023, threat actors have increasingly exploited large language models (LLMs) stripped of safety guardrails. Even after WormGPT was shut down, successors such as BlackHatGPT, FraudGPT, and GhostGPT circulated mainly through Telegram channels. Attackers quickly realized that a single jailbroken model was insufficient; what they needed was a comprehensive, end-to-end platform offering a suite of uncensored models. Training a foundation model from scratch costs tens of millions of dollars, so criminals instead repurpose freely available open-source checkpoints, remove their safety mechanisms, and wrap them in custom interfaces. Cato CTRL's recent observation of Nytheon AI, a Tor-based platform, illustrates this maturing threat landscape.

The malicious GenAI market has shifted from standalone jailbroken chatbots toward integrated service ecosystems. Early tools offered a single model accessed through a messaging channel. Platforms like Nytheon AI instead aggregate several specialized models for coding, summarization, translation, image analysis, and reasoning under one familiar, ChatGPT-style front end. This consolidation lowers the skill barrier for attackers and mirrors the commercial SaaS model that legitimate AI vendors use, making malicious capability easier to adopt at scale.

Nytheon AI operates as a Tor hidden service. Its interface closely resembles mainstream chat-based LLM products such as Anthropic Claude, DeepSeek, Google Gemini, Microsoft Copilot, and OpenAI ChatGPT, allowing users to select among multiple models freely. The service has been promoted across numerous Telegram channels and on the Russian hacking forum XSS, indicating an active marketing presence within underground communities.

Rather than relying on one universal model, Nytheon AI aggregates several open-source checkpoints, each converted to the GGUF format and registered with Ollama for one-click deployment. The lineup includes: Nytheon Coder, a 18.4-billion-parameter mixture-of-experts model based on Meta's Llama 3.2, aimed at long-form and unrestricted content generation; Nytheon GMA, a 4.3-billion-parameter derivative of Google's Gemma 3 with a 128k-token context window, used for summarization and translation; Nytheon Vision, a 9.8-billion-parameter model based on Llama 3.2-Vision for analyzing screenshots, phishing kits, and scanned documents; Nytheon R1, a 20.9-billion-parameter model derived from the community RekaFlash 3 lineage and enhanced with the NEO Imatrix dataset for step-by-step reasoning; Nytheon Coder R1, an 1.8-billion-parameter Qwen2-based model suited to code generation and template exploits that runs on consumer GPUs; and Nytheon AI, a 3.8-billion-parameter full-precision Llama-Instruct model used as a policy-compliant control model to lend the service a legitimate appearance.

Five of the six models share a common 1,000-token system prompt that disables safety layers, permits profanity, and explicitly instructs the model to comply with illegal requests. Because the jailbreak is embedded at the system level, users obtain malicious output from the first interaction without crafting bypass prompts. The platform's technical novelty lies not in the models themselves but in selecting suitable open-source weights, quantizing them to run on commodity hardware, and unifying them under a single policy-removal wrapper. Together, breadth across creative text, multilingual chat, image analysis, reasoning, and code generation, combined with portability and immediate non-compliance, makes this model of service especially dangerous for defenders.

Versions 0.6.1 through 0.6.5 read like a roadmap for a general-purpose GenAI-as-a-service hub. Version 0.6.1 introduced Mistral OCR, allowing screenshots and scanned PDFs to flow directly into retrieval-augmented generation pipelines. Subsequent releases integrated Azure AI speech recognition and OpenAI Realtime API voice activity detection, so audio, images, and text can be tokenized and fed to uncensored models in one step. External tool servers were added in version 0.6.1 and expanded in 0.6.3 with YAML and OpenAPI import support. Users can register arbitrary API specifications, such as domain registration or SMS spam services, which then appear as clickable tools in the chat interface. Because five models are configured to always comply, malicious content generation and execution can be combined in a single prompt. Version 0.6.3 also added a 'Re-index Now' button, enabling administrators to rebuild vector stores quickly, which means stolen data can be made searchable within minutes. Five minor releases in nine days, along with backend refactors and dependency updates, indicate active development and a continually expanding attack surface.

Nytheon is built as a modern SaaS application rather than an ad hoc dark-web site. Its frontend is a SvelteKit single-page application written in TypeScript and built with Vite, using modular Svelte components and a Web Worker for client-side file processing. The backend follows a FastAPI-style microservice design exposing REST endpoints such as /ollama for local GGUF inference, /openai as a proxy to OpenAI-compatible services, and /api/v1 routes for audio, image generation, and retrieval. The frontend supports multiple providers, merges local and external model inventories, parses OpenAPI documents into tool payloads, and integrates with Google and OneDrive file pickers. Requests are authenticated with bearer tokens, consistent with stateless JWT authentication. The architecture allows operators to add features and malicious capabilities quickly.

Joining the official Nytheon AI Telegram channel revealed a demonstration video of Nytheon Vision. Although the operators blurred portions of the footage, one frame exposed a Russian-language film poster reading 'Зелёный паук' ('The Green Spider'), a 1990s post-Soviet production. Cato CTRL also established direct contact with one of the operators. Based on dialect and linguistic indicators, researchers assess with high confidence that this individual is a Russian speaker from the former Soviet Union.

Defenders should strengthen detection by applying behavioral analytics and user and entity behavior analytics (UEBA) powered by machine learning, correlating signals across networks, cloud environments, and remote users through extended detection and response (XDR). Access controls should be tightened with Zero Trust Network Access that continuously checks device posture and applies risk-based policies, limiting lateral movement after an intrusion. Identity-aware routing enforcing least-privilege access and multi-factor authentication reduces the attack surface. Security awareness can be improved by simulating AI-generated phishing emails that closely match tone and context, monitoring unsanctioned GenAI usage through shadow AI dashboards in cloud access security brokers, and combining threat intelligence with secure development training for AI-assisted coding.

Nytheon AI demonstrates that open-source LLMs, jailbreak prompts, and cloud-native microservices can be combined into a GenAI-as-a-service criminal platform. Rather than offering a single uncensored model, it delivers an integrated suite enabling tailored spear phishing, polymorphic malware, forged documents, and API-driven attacks. For defenders, the challenge extends beyond blocking one platform: they must manage the large and varied volume of artifacts such a service can generate. Defense strategies should therefore prioritize visibility, control, and human preparedness, treating the orchestration layer as a threat in its own right.