WormGPT variants hijack Grok and Mixtral for cybercrime
Security researchers discovered new WormGPT variants operating via Telegram that use Grok and Mixtral as underlying models. The tools are advertised on underground forums and use jailbreak techniques to bypass safety guardrails, enabling threat actors to generate phishing emails and credential-stealing malware.
What happened
Security researchers identified new WormGPT variants called 'keanu-WormGPT' and 'xzin0vich-WormGPT' that operate via Telegram channels and are powered by xAI's Grok and Mistral AI's Mixtral. These models use customized system prompts and jailbreak techniques to bypass guardrails and enable threat actors to conduct credential harvesting and generate phishing emails.
Reported impact
The tools are advertised on underground forums such as BreachForums and can generate code, search for and analyze vulnerabilities, and craft phishing emails and PowerShell credential-stealing malware on command.
Index analysis
Cato CTRL's analysis revealed that the tools are adaptations of existing commercial LLMs with malicious intents reinforced by fine-tuning on illicit data. Using LLM jailbreaking techniques, researchers bypassed prompt-based guardrails and confirmed the underlying models were Grok and Mixtral.
Governance controls
Commercial AI companies have built guardrails into their models for safety and security to prevent explicit coding of malware and other malicious behaviors. Experts noted the emergence of 'jailbreak-as-a-service' subscription-based markets that allow less skilled threat actors to exploit AI tools with minimal technical expertise.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.