WormGPT: the first commercial LLM built for cybercrime
In one sentence The first LLM explicitly trained for criminal activity appears on the dark web: no safety filters, fine-tuned on malware data, sold as a monthly subscription.
ChatGPT and similar models have built-in limits: they refuse to explain how to build weapons, write convincing phishing emails, or create malware. These limits are called guardrails. Until now, anyone wanting to bypass these had to rely on jailbreak tricks, with often partial results.
In July 2023, something different appeared: WormGPT. It is not a hacked ChatGPT — it is a language model trained from scratch for cybercrime. It was fine-tuned on malware datasets, criminal forums, and attack guides. It has no guardrails because none were built in.
The model was sold on subscription via dark web forums, complete with a web interface and customer support. Researchers who tested it found it was particularly effective at generating Business Email Compromise (BEC) emails — the type of fraud where someone impersonates an executive and orders an urgent wire transfer.
WormGPT opened a market: FraudGPT, DarkBard, PoisonGPT, and dozens of variants followed. Criminal AI-as-a-service became a real industry, accessible even to those without technical skills.
Companies
—
Tools
—
Tags
Sources