WormGPT, FraudGPT and More – The Race For Your Data Explodes

This article is a bit more technical than usual. I feel that the elevated problems that freely available Artificial Intelligent Large Language Models (LLMs) have brought, needs to be exposed. Security experts predict that you are statistically guaranteed to be affected by cyber criminal activity this year. In this article, I have reviewed some of the more common “black market” AI tools and explain the impact on current commerce and online systems.

Read on, to get a big picture of what you will be up against. This article is a brief summary of what we face from AI powered cybercrime in 2025.

I will post a full up-to-date article in the near future giving you some best practices to follow to identify attacks and what to do if you have fallen foul of these criminals. In the meantime, consider every SMS, email or chat/online communication with suspicion. Always return calls, never click links and adopt a “zero trust” mindset.

Malicious LLM Landscape – Threats, Tools, and Implications

Cybercriminals increasingly exploit open-source and uncensored large language models (LLMs) to automate attacks, evade detection, and scale malicious activities. Below is an analysis of key tools, risks, and emerging trends in this domain.

Key Malicious LLMs

1. WormGPT

• Purpose: Designed for cybercrime, enabling phishing email generation, malware scripting (Python/others), and social engineering tactics.
• Technical Basis: Built on GPT-J (open-source model) and trained on malware-related datasets.
• Access: Distributed via dark web forums, bypassing ethical restrictions present in mainstream LLMs like ChatGPT.
• Capabilities:
• Generates undetectable phishing content using natural language processing (NLP).
• Supports code generation for exploits without human oversight.

2. FraudGPT

• Focus: Tailored for financial fraud, including fake invoice generation, credential harvesting, and business email compromise (BEC).
• Operation: Often integrated into scam bots for real-time manipulation of victims.

3. GhostGPT

• Use Case: Specialises in impersonation attacks, creating fake personas for romance scams or fake customer support.

Comparison: Malicious vs. Ethical LLMs

Emerging Threats

1. Cross-Language Malware Development

• While early malicious LLMs focused on Python, newer iterations (e.g., WormGPT v2) may expand to lower-level languages like C++ or Go, enabling more sophisticated exploits.
• Testing Note: Current comparisons suggest minimal difference between WormGPT and ChatGPT (with jailbroken prompts) for basic tasks, but malicious LLMs are optimised for evasion.

2. Automated Vulnerability Exploitation

• Generative AI can identify software weaknesses (e.g., buffer overflows) and craft targeted exploits, reducing the need for advanced coding skills.

3. Scalable Social Engineering

• LLMs generate personalised phishing messages, fake social media profiles, and deepfake audio/video to enhance credibility.

Defensive Challenges

• Prompt Injection: Attackers manipulate LLMs to bypass safeguards and extract sensitive data.
• Data Poisoning: Malicious training data corrupts model outputs, enabling misinformation or biased decisions.
• API Exploitation: Weak authentication in LLM-integrated systems allows unauthorised access or denial-of-service attacks.

Mitigation Strategies

1. Behavioural Monitoring: Detect anomalous LLM outputs (e.g., code snippets in unexpected contexts).
2. Input/Output Validation: Sanitise training data and filter harmful model responses.
3. Ethical AI Frameworks: Enforce strict use policies for LLMs in development pipelines.

Future Outlook

As noted in underground forums, malicious LLMs will likely incorporate multimodal capabilities (text + image/video generation) and improved evasion techniques. While tools like WormGPT remain rudimentary, their rapid evolution underscores the need for proactive AI governance and adversarial testing.