News
Shared Prompts and Fraud Kits Lower Barriers to Cybercrime
Cybercriminals don't necessarily need to build their own malicious AI models. Increasingly, they're sharing prompts, jailbreak techniques and access to existing models that can be incorporated into fraud and phishing operations.
That's one of the trends emerging from Flashpoint's latest AI Threat Report, which examined how artificial intelligence was discussed and used across illicit forums, marketplaces and chat services during April 2026. The threat intelligence company's analysts identified 2,328,958, or 2.3 million, posts discussing AI in the context of illicit activity during the month.
The activity covered everything from identity verification bypass and fraud scripting to synthetic-media impersonation and jailbreak techniques. But Flashpoint found discussions about developing custom malicious LLMs appeared less frequently than conversations about making existing AI tools easier and more reliable to use
Instead, users exchanged prompts, requested jailbreak methods, and sought alternative models perceived as having fewer safeguards or moderation controls. Users frequently reposted and refined the same techniques as they discovered what worked.
AI Fraud Becomes Easier to Copy
Telegram alone accounted for nearly 1.4 million AI-related posts Flashpoint observed in April. The cloud-based messaging and social media platform served as a major distribution hub for prompts, jailbreak techniques, fraud tools, and service advertisements.
Discord communities tended to focus on individual models and prompt collections, while GitHub Gist and paste sites were used to distribute scripts and supporting material.
Methods also moved between those communities. The security company observed jailbreak prompts being reposted with revised wording, while phishing workflows moved from forums to paste sites and then onto Telegram.
Users tested the techniques, requested changes, and redistributed updated versions when platform restrictions changed or existing methods stopped working.
That reuse could lower the technical barrier for AI-enabled fraud. Rather than inventing a new attack, users can adopt techniques others have already tested and adapt them to their own targets.
Identity fraud was particularly prominent. Flashpoint counted 63,763 posts advertising or discussing AI-assisted know-your-customer, or KYC, bypass methods during April.
Those offerings included synthetic video designed to imitate live verification, cloned voices, scripted interactions and bundled "KYC bypass kits."
Some combined synthetic video with fake documents and AI-generated scripts intended to support an impersonation attempt from beginning to end.
Criminals Look for Fewer AI Guardrails
Flashpoint also observed a significant increase in discussion around VeniceAI, which it attributed partly to new Reddit and Discord communities dedicated to the platform.
The interest reflects a broader search for models that users believe impose fewer restrictions than mainstream services such as ChatGPT or Gemini.
Flashpoint researchers found users sharing jailbreak prompts, phishing and fraud prompt collections, step-by-step instructions, and prompts tailored for impersonation and social engineering.
The threat intelligence company argues that prompt engineering is effectively developing into its own service layer within illicit communities.
The value isn't necessarily in owning a specialized malicious model. It's in finding a repeatable way to make available models produce useful results for fraud, phishing and impersonation.
The potential consequences are already substantial. The cyber threat intelligence company opens its report with an example of a finance employee who transferred $25 million after joining a video call where every other participant was AI-generated.
Flashpoint's April data suggests the techniques behind attacks like that are becoming easier to distribute and reproduce.
"The emphasis remains on accessibility, portability, and ease of use rather than custom model development," Flashpoint researchers wrote.