Generative AI is often placed into the most extreme risk scenarios: a highly autonomous AI agent breaks free of control, connects to the internet, uses hacking tools, and ultimately takes over financial systems, computing resources, and even critical infrastructure. However, in a recent paper, researchers from the Cambridge Cybercrime Centre at the University of Cambridge, the University of Edinburgh, and the University of Strathclyde argue that if we want to understand the real threat of AI to cybercrime, these science-fiction-style imaginings may actually miss the point.
Underground forums are more interested in ChatGPT—far beyond Dark AI
The paper, titled “Stand-Alone Complex or Vibercrime? Exploring the adoption and innovation of GenAI tools, coding assistants, and agents within cybercrime ecosystems,” is written by Jack Hughes, Ben Collier, and Daniel R. Thomas. It was submitted to arXiv on March 31, 2026.
The authors argue that the impact of generative AI on cybercrime should not be understood only from the question of whether AI can write malware; instead, underground cybercrime markets should be viewed as an ecosystem made up of tool vendors, service providers, low-skill operators, and small-time criminal entrepreneurs.
The paper introduces two concepts as the upper and lower bounds of AI’s impact on cybercrime. The high-end scenario is called Stand-Alone Complex—essentially “crime-gang-in-a-box”: a mature AI agent packages what used to require a multi-person division of labor in cybercrime-as-a-service into a semi-automated system, enabling a single actor to execute processes that previously required a criminal team.
The low-end scenario is called Vibercrime, referring to vibe coding, coding assistants, and chatbots lowering parts of the technical barrier, but without truly reshaping the business model and economic structure of cybercrime.
The research team’s conclusion is quite counterintuitive: so far, underground cybercrime communities are indeed highly interested in tools such as ChatGPT, Claude, Gemini, Cursor, Copilot, and WormGPT, but there is no evidence that generative AI has already massively disrupted the cybercrime ecosystem. The paper points out that AI currently functions more like a productivity tool being absorbed into existing criminal workflows, rather than creating an entirely new criminal industrial revolution.
Back from “hacker blockbuster” to the underground economy: cybercrime already looks like a set of small tech companies
The paper first reviews the evolution of the cybercrime ecosystem. In the early days, cybercrime resembled an experimental culture of a few high-skill hackers, emphasizing technical mastery, anti-authoritarian attitudes, and creativity. But after the 2000s, cybercrime gradually became industrialized—turning into a market with division of labor across tools, scripts, templates, initial access permissions, botnet rentals, and customer support—what is commonly known as cybercrime-as-a-service.
The authors argue that underground cybercrime markets rarely invent cutting-edge technology on their own. Real vulnerability research, advanced red-team techniques, and new attack methods mostly come from academic research, cybersecurity companies, or government security organizations; underground criminals are better at repackaging mature technology, copying tools from legitimate industries, developing business models, and automating boring but profitable workflows.
This is why the authors think AI’s real impact on crime may not be that “newbies suddenly learn 0-days,” but rather in more mundane areas: automated customer service, generating scam content, translating scripts, managing accounts, handling back-office processes, optimizing SEO fraud, running community bots, or making already highly automated low-margin gray-market operations more efficient.
Research method: tracking 15 years, over 100 million pieces of underground forum and chat data
The importance of this paper is that it does not rely solely on lab tests, nor on inferring from a few cybersecurity-company case studies. Instead, it uses the CrimeBB dataset from the Cambridge Cybercrime Centre. The dataset covers more than 15 years and more than 100 million posts from underground forums and chat channels, including topics such as account takeovers, SEO scams, game cheating, passive income, romance scams, and more.
The research team used keyword searches with artificial intelligence, LLM, GPT, Claude, Gemini, prompt, Copilot, vibe coding, OpenAI, model, generative, machine learning, AI, and so on. They initially obtained 808,526 threads; then, after excluding discussions before ChatGPT was released, they focused on data between November 1, 2022 and December 10, 2025, finally arriving at 97,895 threads for analysis.
The authors further combined topic modeling, keyword tracking, LLM-assisted categorization, and manual qualitative analysis. Notably, the research team also admitted that when they used a local LLM to categorize underground forum discussions, they found the LLM was not reliable for fine-grained classification. About 80% of the posts that the model labeled as relevant were indeed related to AI or vibe coding, but the specific categories were almost always wrong.
This, ironically, becomes an interesting supporting clue in the paper: LLMs often can only help researchers find “things they already know to ask about,” and still have clear limitations as exploration tools.
Underground forums talk about ChatGPT the most; WormGPT is less important than expected
From keyword trend data, ChatGPT is the AI product most frequently discussed in underground forums. Discussions of Claude show steady growth; after the release of Gemini 1.5, Gemini shows a clear increase. Grok appears in a few short bursts of discussion. By comparison, Codex is discussed less, and while jailbroken models like WormGPT draw heavy attention from cybersecurity media, their discussion volumes in forums do not form a sustained breakout.
The paper notes that underground communities do have cultural excitement around so-called Dark AI. Ads appear for WormGPT, dark versions of ChatGPT, unrestricted models, and offensive AI tools, and many people ask how to access them for free. But the research team found that these discussions mostly stay at “how to get the tools,” “imagining how AI will change the hacker world,” or “testing whether models will answer illegal questions,” rather than successfully using these tools to develop criminal capabilities at scale.
More importantly, the research team saw no clear evidence that newcomers can learn truly viable hacking skills from Dark AI, or generate stable, working malware tools. Instead, some forum users complain that code produced by these tools is unreliable and requires a lot of professional knowledge to fix. This makes jailbroken LLMs more like a kind of underground cultural performance than a major technical breakthrough in the cybercrime ecosystem.
AI didn’t turn newbies into hackers—it’s more like replacing Stack Overflow and a cheatsheet
One of the most important findings in the paper is that AI does not significantly lower the core skill threshold of cybercrime. For users who already have the capability, coding assistants can help write small chunks of code, debug, fill in syntax, and handle general software engineering tasks. But this is more akin to replacing Stack Overflow, cheatsheets, Google searches for error messages, and copy-paste code snippets, rather than creating entirely new criminal capabilities.
For low-skill users, the effectiveness of vibe coding is limited. The reason is simple: they may not know whether the code generated by AI will actually work, and they do not understand how to integrate, fix, or maintain it. Rather than starting from scratch with a chatbot to build an unstable tool, many forum newbies still prefer using existing scripts, templates, tutorial packs, or tools that someone else has already prepared.
In other words, AI has not upgraded “script kiddies” directly into advanced hackers; it is more like improving the efficiency of people who already know how to code a bit more. This also explains why the authors think that even if people worry about the rise of “Vibercriminal,” they may be overestimating the magnitude of the current changes.
What AI truly changes is SEO scams, community bots, and romance scams
Although the paper counters the panic narrative of an AI-driven crime explosion, it does not mean AI has no criminal uses. The research finds that the most obvious adoption scenarios for AI are existing large-scale, low-margin, highly automated gray-market activities, such as SEO scams, community bots, automated AI article generation, content farms, parts of romance scams, and social engineering.
These scenarios share a common feature: they already heavily depend on producing large volumes of content, large numbers of accounts, lots of repetitive work, and exploiting platform rule loopholes. Generative AI can improve copywriting quality, enhance translation ability, vary the patterns of junk content to evade simple rule-based detection, and make already existing automated workflows cheaper and easier to scale.
Therefore, the cybercrime risk brought by AI may not be “an AI agent launching a hacker war by itself,” but rather a more realistic, more boring, and closer-to-the-core-of-platform-economics problem: it allows existing gray-market operations to scale more easily across content, accounts, advertising, SEO, community manipulation, and low-level scams.
Stand-Alone Complex hasn’t appeared yet, but platform AI could create new attack surfaces
For the high-end scenario Stand-Alone Complex, the authors believe there is currently no evidence that it has fully formed. AI agents have not yet integrated ransomware, DDoS, botnet management, payments, customer support, and infrastructure operations into a genuinely “criminal team boxed product.” But the paper also does not completely rule out this future.
The authors caution that if online platforms themselves shift from the past 20 years of display ads, search, and social traffic models toward a new architecture centered on chatbots and AI-generated answers, gray-market players familiar with SEO, content farms, account farms, bot infrastructure, and the offense-defense game around platform rules may find new opportunities to profit. Put differently, AI may not make underground criminals more high-tech, but it could change the economic structure of legitimate platforms—thereby changing where criminals can arbitrage.
This is also the most notable extension point in the paper: the biggest impact of AI on cybercrime may not be underground forums inventing some new technology themselves, but rather the legitimate AI industry changing the entire internet’s traffic, content, advertising, search, and automation structures—helping existing gray-market actors find new gaps.
Author’s suggestion: don’t panic, but don’t ignore the amplifying effect of low-level automated crime
In the paper’s final recommendations to policymakers, the industry, and law enforcement, the advice can be condensed into a single sentence: don’t panic. The research team believes that, so far, AI tool adoption in underground cybercrime ecosystems remains fragmented, incremental, and non-revolutionary; you cannot directly apply AI coding adoption patterns from the legitimate software industry to cybercrime business because many criminal business models do not actually rely on advanced technical capabilities.
But “don’t panic” does not mean “no risk.” The authors point out that model guardrails, fine-tuning, and friction in usage are still effective—especially in low-level, high-volume, and heavily automated misuse scenarios. These can increase costs for criminals and limit the scale of gray-market activity through saturation and competition for resources. These measures cannot stop motivated high-end attackers, but they can slow the expansion enabled by low-cost misuse.
This paper provides a calmer, more realistic framework than “AI hacker apocalypse.” Generative AI has not turned newbies into hackers overnight, and it has not built fully automated criminal organizations. Instead, it is more like bringing underground criminals into the era of AI-assisted work. The real issue is not whether AI will make criminals superheroes, but how it will amplify the marginal returns of existing gray-market operations, platform arbitrage, content automation, and low-cost scams.
This article says AI didn’t turn newbies into hackers! UK researchers: AI is mostly used for spam content and emotional-romance scams First appeared on ChainNews ABMedia.
