A criminal AI service has been observed selling access to frontier-model capabilities for $12.99 a month, but ThreatDown’s analysis indicates Kriminal is not operating a proprietary criminal model. Instead, its production code points to legitimate providers unwittingly supplying much of the technology underneath the service.
ThreatDown said Aug. 18 that Kriminal’s code identifies xAI’s Grok as its primary inference engine, while OpenRouter routes specialist models including Mistral Large and Llama 3.3.
Anthropic’s Claude is listed for long-context work, with Google Cloud, Cloudflare, Tavily and NowPayments also appearing in the wider stack. What Kriminal adds is a jailbreak layer: a prompt-injection block appended to every request, instructing the underlying models to ignore any limits on their output.
The platform offers plans ranging from free access to $99 a month. Its $12.99 AGENT tier includes about 200 messages, while the service also markets capabilities including OSINT analysis, on-chain tracing and offensive coding.
Aviv Nahum, Co-founder and CEO at Above Security, told CISO Voice that there may be much less “criminal AI” underneath the branding than Kriminal suggests.
“Criminals are doing what software companies have always done,” Nahum said, describing a model in which powerful technology built by someone else is repackaged for a specific customer. “The underlying capability is becoming a commodity.”
Model safeguards are not a complete security strategy
Diana Kelley, CISO at Noma Security, told CISO Voice that the development could be a bellwether for a broader shift in cyber offense as advanced capabilities become cheaper and more accessible.
“As advanced offensive capability becomes cheaper and more accessible with AI, attackers can find and exploit weaknesses at a speed and scale that tilt the economics of cybercrime in their favor,” Kelley said.
Her recommendation is to use AI defensively to uncover exposure and reduce accumulated security debt before attackers can weaponize it. “CISOs need to fight fire with fire,” she said.
Ram Varadarajan, CEO at Acalvio, told CISO Voice that the price point itself is significant, with frontier-level capabilities for cybercrime now available for $12.99 a month.
Varadarajan said model-level filtering should not be treated as an impenetrable final barrier. “Behavior is harder to disguise than content,” he said, arguing that security teams should focus on what AI-enabled systems actually do.
The Kriminal case also illustrates the visibility problem created by distributed infrastructure. An AI provider may see activity reaching its models, a cloud or network provider may see another part of the stack, and a payment provider may see the transaction, without any one company necessarily seeing the complete operation.
For CISOs, the practical implication is straightforward: legitimate AI providers’ safeguards can reduce abuse, but they cannot be the organization’s last line of defense once capable models are available through multiple routes.