Cybercrime

Cybersecurity researchers say cybercriminals are turning to a growing underground market of customised large language models designed to support low-level hacking tasks.

A new report from Palo Alto Networks’ Unit 42 describes how dark web forums promote jailbroken, open-source and bespoke AI models as hacking assistants or dual-use penetration testing tools, often sold via monthly or annual subscriptions.

Some appear to be repurposed commercial models trained on malware datasets and maintained by active online communities.

These models help users scan for vulnerabilities, write scripts, encrypt or exfiltrate data and generate exploit or phishing code, tasks that can support both attackers and defenders.

Unit 42’s Andy Piazza compared them to earlier dual-use tools, such as Metasploit and Cobalt Strike, which were developed for security testing but are now widely abused by criminal groups. He warned that AI now plays a similar role, lowering the expertise needed to launch attacks.

One example is a new version of WormGPT, a jailbroken LLM that resurfaced on underground forums in September after first appearing in 2023.

The updated ‘WormGPT 4’ is marketed as an unrestricted hacking assistant, with lifetime access reportedly starting at around $220 and an option to buy the complete source code. Researchers say it signals a shift from simple jailbreaks to commercialised, specialised tools that train AI for cybercrime.

Another model, KawaiiGPT, is available for free on GitHub and brands itself as a playful ‘cyber pentesting’ companion while generating malicious content.

Unit 42 calls it an entry-level but effective malicious LLM, with a casual, friendly style that masks its purpose. Around 500 contributors support and update the project, making it easier for non-experts to use.

Piazza noted that internal tests suggest much of the malware generated by these tools remains detectable and less advanced than code seen in some recent AI-assisted campaigns. The wider concern, he said, is that such models make hacking more accessible by translating technical knowledge into simple prompts.

Users no longer need to know jargon like ‘lateral movement’ and can instead ask everyday questions, such as how to find other systems on a network, and receive ready-made scripts.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!