What is the dark LLM economy and how is it powering modern-day scams 

By Sathyan Sethumadhavan, Senior Director – AI, UST

The “dark LLM economy” is changing the world of cybercrime in a big way. Uncensored, jailbroken large language models are no longer just cool things that people do in secret. The models are the building blocks of a growing, industrialised fraud ecosystem. 

Generative AI has turned cybercrime from a manual, skill-intensive job into a very useful service.  The work of Harvard’s Berkman Klein Center adds an important lens of interpretability and transparency. Researchers show in Inside the Black Box that AI systems can guess sensitive information about users and change outputs in ways that users can’t see. This is important for fraud defense because hidden inference logic can be used to manipulate, discriminate, and target people with precision. Harvard’s larger 2025 Action Report also makes it clear that trust, transparency, and open governance tools are no longer “ethics extras” but are now necessary for business.

Threat actors have made it much easier for people to get into the business by using AI throughout the entire attack lifecycle, from reconnaissance to post-compromise execution. They have also achieved an unprecedented operational scale. For leaders of businesses, this is an urgent sign to see that old security perimeters are failing, and they need to switch to identity-centric, behavioural defense.

Industrialisation of cybercrime

The rise of dark LLMs has changed the economics of digital fraud in a big way. For a small upfront cost, new threat actors can get uncensored AI and make a lot of money from it. This has led to the rise of Fraud-as-a-Service (FaaS), which combines AI with modular crime platforms.

A solitary scammer can now operate with the velocity of an entire syndicate, generating malicious code and thousands of unique phishing lures instantly. INTERPOL’s 2026 Global Financial Fraud Threat Assessment warns that AI-enhanced fraud yields significantly greater profitability, tying operations to broader polycriminal networks. The March 2026 disruption of the Tycoon2FA ecosystem, which included the seizure of 330 domains, highlights the formidable maturity and cross-border coordination of these AI-powered supply chains.

Evolution of social engineering

There is no longer a time when phishing emails were poorly written and easy to spot. Cybercriminals now use dark LLMs to scrape hacked databases and business profiles, creating hyper-personalised lures that perfectly copy the way trusted executives and vendors talk.

The threat goes way beyond than just writing convincing emails. Attackers use self-driving AI agents that attack victims at the same time on many channels, keeping the same identities and changing their behavior in real time based on how victims respond. According to recent Microsoft threat intelligence, modern campaigns use dynamic code generation and on-click authentication flows to run active operations that poll backend states to steal more effectively. Also, the combination of real-time voice cloning and automated translation breaks down language barriers, making it possible for fraudsters to make fake calls that sound like they’re coming from familiar authorities and to spread scams that used to be limited to one area.

Collapse of legacy defenses

Old-fashioned anti-fraud programs were made for a different kind of threat, one that relied on obvious problems, stable signatures, and attackers who acted in predictable ways. This model has completely failed. Dark LLMs make content that is grammatically and contextually perfect, which makes content-based filters and signature-matching systems useless.

AI doesn’t send out static payloads instead, it makes polymorphic attacks that send out thousands of different messages with the same bad intent but different wording. Also, threshold-based alerts quickly become too much for AI-generated interactions to handle because they happen so quickly. Cybercriminals have gone from simple spoofing to dynamic identity compromise, which lets them easily get around security measures like Multi-Factor Authentication (MFA) by stealing session artifacts. Rule-based defenses that are too strict don’t work anymore against AI-driven deceptions that can change quickly.

Strategic pivot to identity and behaviour 

As AI-generated deception makes content inspection useless, organisations need to change their strategic focus to identity assurance. The key question for security teams is no longer “Does this look suspicious?” but “Can we cryptographically prove who started this action?”

Organisations must abandon single-point authentication, assuming compromise and enforcing continuous, stepped-up verification for sensitive actions. While AI can mimic a CFO’s writing, it cannot replicate human physics. Robust device intelligence and behavioural analytics measuring typing cadence and navigation sequences are essential alternative detection surfaces. Combating industrialised deception requires breaking down internal silos. Fraud, Security Operations Center (SOC), and trust-and-safety teams must operate from a shared risk graph, combining identity, device, and post-authentication telemetry.

Securing the internal AI attack surface

Business leaders have to deal with two big risks at the same time. On one hand, they need to protect themselves from outside attacks by bad actors using advanced language models. On the other hand, they must make sure that the AI systems they use inside their companies don’t become vulnerable to attacks. If a company’s language model is not properly secured, it can be manipulated by injecting malicious prompts. This can have serious consequences, such as stealing sensitive information, executing unauthorised commands, or even moving freely within the company’s network.

Companies should handle their internal AI systems with the same level of caution they use for their most privileged employees. This means keeping a list of approved AI models, having strict controls over how these models interact with other systems, and making sure the information going in and out of these models is thoroughly checked and cleaned. It’s especially important for tasks that involve high risks, like handling payments or updating personal information, to have a human reviewer double-check the work. As old ways of dealing with threats become less effective, training employees should include a requirement for independent verification of all critical messages to ensure everything is legitimate.

Conclusion:

Global institutions, including the World Economic Forum and INTERPOL, confirm that dark LLM-enabled fraud is not a fleeting trend. It is an AI-driven organised crime. The commercialisation of these illegal tools has made traditional fraud methods quicker to launch, cheaper to execute, and endlessly scalable.

Executives should aim not for zero fraud but for significantly reducing the return on investment for attackers and their time inside the system. By treating generative AI fraud as a key business risk, speeding up identity defenses, and carefully managing internal automation, organisations can create lasting strength against the constant threat of the dark LLM economy.