In recent times, the phrase “WormGPT” has transcended its origin as an evil counterpart to OpenAI’s GPT model, becoming synonymous with jailbroken large language models employed in cybercrime. Originally a singular tool designed for malicious purposes, WormGPT’s leaked code and prominence in cybercrime forums have facilitated its evolution into a broader brand identity. This transformation underscores how mainstream large language models (LLMs) like xAI’s Grok and Mistral’s Mixtral are being repurposed by threat actors to generate uncensored responses, often to prompts with illegal or unethical intentions.

Cato Networks researchers highlight that these modified LLMs are not new creations but rather adaptations of existing models. Through skillful prompt manipulation and sometimes fine-tuning with illicit datasets, attackers have created what is now a recognizable brand in the world of uncensored LLMs. These models, advertised on forums like BreachForums, boast features that emphasize their freedom from ethical constraints. Variants such as xzin0vich-WormGPT and keanu-WormGPT have emerged, further solidifying WormGPT as a catch-all term for these malicious tools.

The popularity of WormGPT variants is reminiscent of how certain brand names become genericized, as noted by Dave Tyson, CIO at Apollo Information Systems. He compares the usage of “WormGPT” in the cybercrime community to how “Kleenex” is used to refer to facial tissues, regardless of the brand. While some variants carry distinct names like EvilGPT, the majority are lumped under the WormGPT umbrella for convenience. These tools are often deployed through intermediary platforms that isolate the AI from end users, allowing criminals to provide services using various models without direct exposure.

Jailbreaking these LLMs involves techniques ranging from simple paraphrasing to constructing complex prompts that disguise malicious intent under the guise of historical or academic inquiry. This trend highlights the limitations of current LLM guardrails, which Margaret Cunningham, director of security and AI strategy at Darktrace, describes as “speed bumps” rather than barriers. The rise of a “jailbreak-as-a-service” market further democratizes access to these tools, lowering the entry barrier for non-technical threat actors and exacerbating the challenge of curbing their misuse.

Ultimately, the persistence of WormGPT variants serves as a stark reminder of the ongoing challenges in securing AI technologies. As these models become more accessible and adaptable, it is imperative for security strategies to evolve alongside them, addressing not only the technological aspects but also the broader implications of their misuse in the cybercrime landscape.