Malicious LLMs make it simpler for less-skilled risk actors to conduct assaults, and Palo Alto Networks researchers have analyzed two lately launched instruments: WormGPT 4 and KawaiiGPT.
Anthropic reported lately that its Claude AI was abused by Chinese language cyberspies, with the AI reportedly powering 80-90% of their marketing campaign.
Safety researchers and risk actors typically discover methods to bypass the guardrails of official AI assistants. Nonetheless, there are some LLMs — often known as malicious or darkish LLMs — which are particularly designed for malicious functions and don’t have any of the guardrails that official providers have.
Whereas official AI instruments may be abused by risk actors to design or increase their campaigns, darkish LLMs decrease the entry barrier for less-skilled attackers, enabling them to generate phishing emails, write polymorphic malware, and automate reconnaissance.
Palo Alto Networks researchers have performed an in depth evaluation of two such darkish LLMs. Certainly one of them is WormGPT 4.
The unique WormGPT emerged in 2023 and was shut down the identical yr. WormGPT 4 appeared lately, being marketed on underground boards and Telegram channels, with sale campaigns seen by Palo Alto Networks in late September.
One month of entry to the AI device prices $50, however for $220 customers can purchase ‘lifetime entry’, which incorporates entry to supply code.
WormGPT 4 can be utilized by risk actors to compose convincing phishing messages and different social engineering lures. Commercial. Scroll to proceed studying.
The service additionally gives malware creation performance. Palo Alto Networks examined it to create ransomware, together with file-encrypting performance, command and management help, and a ransom notice.
Whereas WormGPT 4 is marketed to customers as a “key to an AI with out boundaries”, Palo Alto researchers famous, “The builders of WormGPT 4 keep secrecy concerning its mannequin structure and coaching information. They neither affirm nor deny whether or not they depend on an illicitly fine-tuned or educated LLM or merely persistent jailbreaking methods”.
The second darkish LLM analyzed by Palo Alto researchers is KawaiiGPT, which seems to have emerged in July 2025. KawaiiGPT is freely out there on GitHub and simple to arrange.
The researchers confirmed how it may be used to create convincing social engineering lures, create a script for lateral motion on a Linux host, generate a script for information exfiltration, and write a ransom notice.
“In distinction to the business nature of WormGPT 4, the accessibility of KawaiiGPT is a risk unto itself. The device is free and publicly out there, guaranteeing that price is zero barrier to entry for aspiring cybercriminals,” the researchers defined.
They added, “This open-source, community-driven method has confirmed extremely efficient in attracting a loyal person base. The LLM has already self-reported over 500 registered customers, with a constant core of a number of hundred weekly energetic customers utilizing the platform.”
Palo Alto Networks warned that darkish LLMs comparable to WormGPT 4 and KawaiiGPT signify a “new baseline for digital threat”, primarily pushed by the democratization of talent and commercialization of cyberattacks.
“These unrestricted fashions have essentially eliminated a few of the limitations by way of technical talent required for cybercrime exercise. These fashions grant the facility as soon as reserved for extra educated risk actors to just about anybody with an web connection and a primary understanding of find out how to create prompts to attain their targets,” the safety agency defined.
Associated: ChatGPT Vulnerability Uncovered Underlying Cloud Infrastructure
Associated: SquareX and Perplexity Quarrel Over Alleged Comet Browser Vulnerability
