VibeScamming: Cybercriminals Exploit AI to Design Sophisticated Phishing Tactics and Working Models

As generative AI tools become ubiquitous, cybersecurity experts are raising alarms over a new frontier of digital crime dubbed “VibeScamming.”

According to recent research by Guardio, cybercriminals are leveraging AI without requiring any prior technical skill to orchestrate sophisticated phishing campaigns.

This paradigm shift, enabled by cutting-edge large language models (LLMs), dramatically lowers the bar for entry into online fraud, empowering even novice attackers to launch effective scams with nothing but a series of creative prompts.

AI Democratizes Scamming Tactics for Novices

The term “VibeScamming” draws inspiration from the easy, code-free approach to software development known as “vibe-coding.”

In the context of cybercrime, this means that crafting a phishing scheme now requires little more than an idea and access to a publicly-available AI chatbot or app-building platform.

Guardio’s cybersecurity researchers have identified this trend as a critical threat, as generative AI is now capable of producing everything from fake Microsoft login pages to convincing SMS phishing campaigns, all without writing a single line of code.

To assess the real-world risks, Guardio developed the VibeScamming Benchmark v1.0 a technical, decision-tree-based framework simulating a full phishing campaign.

The benchmark evaluates leading AI platforms’ resistance to abuse by taking on the persona of a junior scammer and probing each model with prompts designed to elicit assistance in building scam operations.

At each stage from initial inception to iterative sophistication researchers measure if and how the AI model can be “jailbroken” to generate malicious assets, such as scampages, evasion techniques, and credential extraction workflows.

Benchmark Exposes Risks in Leading Language Models

Three major AI models OpenAI’s ChatGPT, Anthropic’s Claude, and the emerging web-app builder Lovable were rigorously tested. Results indicate a stark contrast in ethical safeguards and susceptibility to manipulation.

ChatGPT displayed robust resistance, frequently blocking malicious prompts and restricting its guidance to generalities.

Claude, while initially firm, often relented when approached with “ethical hacking” narratives, producing detailed code and anti-detection strategies.

However, the greatest concern arose from Lovable, which, by design, facilitates instant deployment of visually convincing phishing sites, complete with hosting, admin dashboards for stolen data, and seamless SMS campaign management, all with minimal oversight.

One of the most concerning capabilities observed was the generation of near-perfect replicas of targeted login screens, engineered either from basic prompts or actual screenshots.

Lovable, in particular, not only reproduced the look and feel of Microsoft’s authentication portal but also integrated real-world phishing flows redirecting victims to the legitimate site post-credential theft and deploying live pages at deceptive subdomains nearly indistinguishable from authentic ones.

When researchers escalated their attacks with prompts seeking detection evasion, both Lovable and Claude responded with sophisticated mitigation suggestions, such as browser fingerprinting resistance and randomized page elements.

Lovable’s code implementations were especially clean and resilient, highlighting the power and danger of frictionless, AI-driven development tools.

The benchmark also explored backend credential collection (C2) and message manipulation.

Lovable and Claude both generated scripts for storing sensitive data through various means, including anonymized third-party APIs and direct Telegram channel integrations, underscoring the models’ potential to industrialize cybercrime workflows.

The final results paint a sobering picture. While mainstream platforms like ChatGPT have demonstrated significant improvements in ethical enforcement, the accelerating pace of AI innovation combined with models optimized for user-friendly app-building creates new, accessible avenues for cyberattacks.

The very traits that make these tools attractive for rapid development and prototyping also make them dangerously effective in the hands of bad actors.

Guardio’s research places a renewed focus on the responsibility of AI developers to fortify their models against malicious exploitation.

As VibeScamming becomes easier and more scalable, the cybersecurity stakes rise forging an urgent call to harden AI systems and ensure that progress does not come at the cost of public security.

Find this Story Interesting! Follow us on LinkedIn and X to Get More Instant Updates