AI Goes on Offense
How large language models are changing the economics and scale of cybercrime
How WormGPT works
While WormGPT is often discussed as a singular tool, it is better understood as part of a broader category of illicit LLMs that circulate on underground marketplaces. These models take advantage of recent advances in generative AI but are repurposed to lower the skill barrier for threat actors and accelerate offensive operations.
WormGPT operates like legitimate LLMs in its ability to generate human-like text, but its purpose and behavior differ in critical ways. These models are typically:
- Trained or fine-tuned without safety constraints, allowing them to respond to prompts that legitimate systems would block.
- Advertised on darknet or criminal forums with claims of bypassing content moderation, supporting malware development, or generating fraud scripts.
- Designed to help attackers scale communication-based attacks, such as business email compromise (BEC), phishing, impersonation, or scam outreach.
- Capable of producing syntactically correct code, including scripts intended to automate reconnaissance or exploit workflows.
Although many claims about WormGPT variants lack transparency, their practical function is consistent: To generate harmful content on demand.
Why WormGPT Is dangerous
1. No guardrails or ethical filters
WormGPT responds to prompts that legitimate AI systems reject – for example, writing malware, crafting spear-phishing emails, or producing impersonation messages that target specific individuals.
2. Scale and automation
Because LLMs generate high-quality text instantly, attackers can produce thousands of tailored messages at once, significantly increasing the volume and success rate of phishing campaigns.
3. Enhanced deception
WormGPT outputs are often more articulate, contextually relevant, and grammatically correct than manually written phishing content. This improvement makes social engineering harder for users to spot.
4. Rapid variant growth
As defenders block known tools, new malicious models arise. This includes the likes of FraudGPT, KawaiiGPT, and model variants built on top of Grok, Mixtral, or other emerging architectures. The illicit ecosystem evolves quickly and unpredictably.
WormGPT vs. legitimate AI models
Although WormGPT may resemble mainstream LLMs in interface or function, the differences are stark:
- Intent: Mainstream models aim to support productivity and learning; malicious LLMs aim to support cybercrime.
- Training data: Illicit models may be trained on stolen data or harmful examples.
- Safety layers: Reputable AI providers employ content filters, monitoring systems, abuse prevention, and regulatory compliance.
- Accountability: Malicious LLMs offer no accountability; no provider oversight, no responsible disclosure program, and no transparency.
These distinctions matter because they define the model’s behavior: WormGPT is engineered to facilitate attacks, not prevent them.
Common criminal use cases
Security teams should understand how WormGPT is typically used in real-world scenarios.
BEC and phishing campaigns
Attackers can prompt WormGPT to write realistic emails in the tone of an executive, vendor, or partner. The model can adjust tone, urgency, and context to match known BEC patterns.
Multilingual social engineering
Generative AI can instantly translate or rewrite messages into different languages, expanding the attacker’s reach without requiring language proficiency.
Malware code assistance
WormGPT can create code snippets, obfuscate scripts, or help revise malware components. While outputs are not guaranteed to be production-ready, they reduce the attacker’s effort substantially.
Reconnaissance and research support
Attackers can prompt illicit LLMs for explanations of vulnerabilities, exploitation steps, or lists of common misconfigurations, all content that legitimate AI systems would block or redact.
Fraud and impersonation chats
Models can simulate dialogue to trick victims into sharing sensitive information, posing as employees, vendors, or support agents.
How security teams can respond
Let’s take a look at how security operations might – to varying degrees – be able to respond to this accelerated AI threat:
1. Strengthen detection and response workflows
Because attackers use AI to increase speed and volume, defenders must rely on visibility across identity, endpoint, network, and cloud environments. AI-generated phishing or BEC patterns demand strong detection tied to user behavior, anomalous activity, and cloud access monitoring.
2. Improve phishing and social engineering resilience
Employee training should include examples of AI-generated attacks, which often differ in tone and linguistic precision from traditional scams.
3. Monitor for illicit AI use cases
Threat intelligence teams should track chatter around new model releases, rebrands, and subscription services in underground markets.
4. Implement layered controls
Filtering, authentication, segmentation, and user and entity behavioral analytics (UEBA) all help limit the blast radius of attacks that begin with AI-generated lures.
How WormGPT signals a shift in cybercrime
WormGPT represents a broader transition toward AI-enabled cybercrime, where attackers gain efficiency, anonymity, and reach. Generative AI tools have lowered the barrier to entry: threat actors no longer need strong writing skills or coding expertise to produce credible attack content.
As more malicious LLMs appear – and as they improve through fine-tuning or stolen model weights – the defensive landscape will continue to evolve. Security teams that understand this shift will be better prepared to respond, communicate risk, and reinforce security programs against fast-moving threats.
The evolution of malicious AI models
WormGPT is not an isolated creation; it represents a broader trend in how threat actors adapt new technologies to their advantage. Early malicious AI tools were unsophisticated, often based on repurposed chatbot frameworks with limited capabilities. Over time, however, attackers discovered that generative models are ideal for producing believable text, code fragments, and interaction patterns. As commercial LLMs grew more advanced, underground actors followed suit, leading to a rapid rise in malicious model variants.
Most of these tools share key characteristics. They are intentionally stripped of safety measures, rely on anonymous or pseudonymous creators, and frequently incorporate modified components from leaked or open models. Some variants rebrand quickly to avoid takedowns or marketplace closures, using new names to retain customer interest. Others are designed as subscription services with tiers for “premium features,” mirroring legitimate SaaS models but applied to fraud, phishing, and malware tooling.
Understanding this trajectory helps security teams anticipate how illicit LLMs may evolve next and why preparing for AI-augmented threats is now a foundational element of modern security strategy.
Frequently asked questions
Related reading
Related fundamentals
AI Threat Detection Fundamentals
Human-in-the-Loop Security Fundamentals
AI Security Compliance Fundamentals
Security Operations Center Fundamentals
Recommended blogs
AI Goes on Offense: How LLMs Are Redefining the Cybercrime Landscape
What the First 24 Hours of a Cyberattack Can Teach You About MDR
Speed, Scale, and Immediate Action With Agentic AI Workflows