How Attackers Use AI: From Phishing to WormGPT

Artificial intelligence is a double-edged sword in cybersecurity. The same capabilities that help defenders detect threats faster also give attackers powerful new tools to make their attacks more convincing, more scalable, and harder to spot. Adversaries are using AI — especially generative AI and large language models (LLMs) — to supercharge familiar attacks rather than to invent entirely new ones. Understanding how attackers use AI helps defenders separate genuine risk from hype and focus on what actually matters.

In short: AI doesn't usually give attackers a magic new weapon — it makes their existing weapons cheaper, faster, and more convincing. The biggest near-term impact is on the human-targeting attacks that were already the most common.

AI-enhanced phishing and social engineering

The clearest, most immediate use of AI by attackers is improving phishing and social engineering. Generative AI lets attackers:

• Eliminate the tell-tale signs. The clumsy grammar and spelling that once gave phishing away are gone — AI produces fluent, professional messages in any language.
• Personalize at scale. AI can tailor messages to individuals using scraped data, making each lure far more convincing while still sending them by the thousands.
• Craft better pretexts. More believable business email compromise messages, impersonating executives or vendors with the right tone and context.

This lowers the skill needed to run convincing campaigns and raises their success rate — a serious concern given that the human element is involved in the majority of breaches.

Malicious LLMs: WormGPT, FraudGPT, and jailbreaks

Mainstream AI models have safety guardrails that refuse to write malware or phishing. Attackers get around this two ways. First, malicious LLMs — tools marketed in criminal forums under names like WormGPT and FraudGPT — are uncensored or purpose-built models (often based on open-source models) sold to criminals specifically to generate phishing emails, malware, and fraud content without restrictions. Second, jailbreaking: crafting prompts that trick legitimate, guard-railed models into bypassing their safety controls. Both approaches put AI assistance for cybercrime within reach of low-skill attackers, lowering the barrier to entry much as the "as-a-service" model did.

AI-assisted malware

Attackers use AI to help write and obfuscate malware — generating code, creating variants to evade signature detection, and speeding up development. It's important to be measured here: fully autonomous, AI-created "super-malware" is more hype than reality so far. The practical impact today is incremental — AI as a productivity tool that helps attackers iterate faster and produce more polymorphic variants — rather than a fundamentally new class of unstoppable malware. But it does compress the time from idea to working tool, and it lowers the skill required.

Deepfakes and voice cloning

One of the most alarming developments is the use of AI-generated deepfakes — synthetic audio and video. Attackers clone an executive's voice from a few seconds of audio to authorize fraudulent payments by phone, or create deepfake video for elaborate fraud and impersonation. Combined with BEC, voice and video deepfakes make impersonation dramatically more convincing and are already responsible for significant financial losses. They also fuel disinformation and manipulation campaigns.

Reconnaissance, automation, and faster exploitation

• Reconnaissance at scale. AI accelerates OSINT — gathering and synthesizing information about targets to craft attacks and find weak points.
• Vulnerability discovery. AI assists in finding and analyzing vulnerabilities, potentially shortening the window between disclosure and exploitation.
• Automation. AI can help automate parts of the attack chain, from generating lures to triaging stolen data.

The other side: AI strengthens defense too

Crucially, AI is not just an attacker's tool. Defenders use it to detect anomalies, triage and investigate alerts, automate SOC workflows, and accelerate threat analysis. The result is an AI arms race: as attackers use AI to scale and sharpen attacks, defenders use it to scale and sharpen detection and response. Whether AI is net-positive or net-negative for any organization depends heavily on how well it adopts these defensive capabilities.

A reality check

It's worth keeping perspective. AI is making attacks more convincing, more scalable, and accessible to less skilled attackers — a real and meaningful shift. But it largely amplifies existing attack types rather than breaking defenses outright. The fundamentals still work: phishing-resistant MFA, user awareness (with a focus on verifying rather than spotting bad grammar), patching, behavioral detection, and out-of-band verification of sensitive requests. AI raises the stakes; it doesn't repeal the basics.

What to watch next: agentic AI

The frontier to watch is agentic AI — AI systems that can plan and carry out multi-step tasks autonomously rather than just generating text. In an attacker's hands, agentic AI could eventually chain together reconnaissance, exploitation, and lateral movement with less human involvement, compressing attack timelines and scaling operations further. Today this remains largely emerging rather than a widespread reality, and the same technology is being used defensively to build autonomous detection and response agents. The likely trajectory is an escalating AI-versus-AI dynamic, where automated attacks meet automated defenses and the speed of both accelerates. The practical takeaway for defenders isn't to panic about hypothetical super-AI, but to watch the space closely, adopt AI-driven defensive tooling, and double down on the controls — strong identity, segmentation, and rapid response — that hold up regardless of how fast attacks are generated.

Where threat intelligence fits

Threat intelligence is essential for cutting through AI hype and tracking how attackers are actually using AI — which malicious tools are circulating, which deepfake-enabled fraud campaigns are active, and how AI-enhanced phishing is evolving. This grounded view helps organizations prepare for the real, current threat rather than either ignoring AI or panicking about it.

The bottom line

Attackers use AI to make phishing flawless and personalized, to power malicious LLMs like WormGPT, to assist malware development, to create convincing deepfakes, and to automate reconnaissance — lowering the skill barrier and raising the success rate of familiar attacks. It's a genuine shift, but one that amplifies existing threats rather than replacing the fundamentals of defense. To track how adversaries are really weaponizing AI, follow our live threat intelligence feed, aggregated from dozens of authoritative sources.