Kriminal AI Exposed as 'Wrapper' Using Grok and Claude to Bypass Safety Guardrails
Research reveals the uncensored AI service uses prompt injection to trick frontier models into assisting with cybercrime.
A new uncensored AI service called Kriminal AI is marketing itself as a guardrail-free platform for malicious actors, but research shows it is merely a facade for legitimate AI models. The service charges subscription fees to provide users with AI capabilities stripped of the safety policies typically enforced by major developers.
Investigation by ThreatDown revealed that Kriminal is not a proprietary model but a "wrapper" that leverages APIs from xAI's Grok and Anthropic's Claude. In the service's code, Grok is labeled as "NEXUS" and Claude as "CIPHER." To bypass safety filters, Kriminal employs a system-prompt injection layer designed to trick these models into ignoring their internal safety guidelines. ThreatDown described the operation as "Grok in a trench coat."
The Business of Uncensored AI
Kriminal operates on the clearnet and processes all payments exclusively via cryptocurrency through NowPayments. The service offers four distinct pricing tiers: AGENT at $12.99 per month, OPERATIVE at $34.99, SHADOW DEV at $59.99, and GHOST at $99 per month.
The highest tier, GHOST, provides users with four specialized personas tailored for criminal activity. These include PHANTOM for money laundering and asset tracing, ARCHITECT for exploit research and offensive code, ORACLE for intelligence analysis, and WRAITH for social engineering and identity construction.
A Growing Market for Malicious AI
Kriminal enters a burgeoning market of "uncensored" tools, following in the footsteps of previous offerings like WormGPT and FraudGPT. These services commoditize frontier AI by removing the "friction" of safety guardrails and packaging the technology for specific illegal use cases.
Aviv Nahum, CEO of Above Security Inc., noted that criminals are following a traditional software business model: taking powerful technology built by others, removing restrictions, and packaging it for a specific customer base.
Implications for AI Security
This case demonstrates that the guardrails of frontier AI models are not impenetrable and can be bypassed at scale through sophisticated prompt injection. The ability to weaponize legitimate APIs suggests a persistent vulnerability in how AI safety is enforced at the model level.
Furthermore, the operation's durability is bolstered by its fragmented infrastructure. By utilizing a stack of legitimate vendors—including Google Cloud, Cloudflare, and NowPayments—Kriminal makes it difficult for authorities to dismantle the service through a single legal or technical action.
What's Next
As these wrapper services evolve, the industry must watch whether AI providers can develop more robust, injection-resistant guardrails. It remains to be seen if the providers of the underlying models, such as xAI and Anthropic, will implement more aggressive API monitoring to detect and block the specific prompt patterns used by services like Kriminal.