Artificial intelligence has ceased to be merely a topic for discussion in academic circles and is now a full-fledged element of cyber warfare. My analysis of data from 2025 and the first half of 2026 shows: AI is deeply integrated into real attacks—from reconnaissance to the final generation of malicious code. These are no longer experiments, but established practice.
AI in Action: From Phishing to LLM Attacks
During this period, I recorded at least 18 APT groups that actively used AI in their operations. One illustrative example is the group GTG-1002, which used language models to create phishing emails. The texts were dynamically adapted to a specific victim, allowing them to successfully bypass standard spam filters. This is a qualitative leap: previously, phishing was "mass" and easily recognizable; now it has become targeted and personalized.
Special attention deserves the cluster TAT26-12, which automated the process of gathering intelligence about employees and corporate infrastructure. But the most alarming development is the emergence of specialized tools such as PromptSpy. This software is designed to attack corporate systems based on large language models (LLMs), extracting system prompts and manipulating AI behavior. Another tool, LAMEHUG, allows generating malicious code with minimal human involvement, sharply lowering the barrier to entry into cybercrime.
Shifting Preferences on Dark Forums
A clear shift in hackers' preferences is observed. OpenAI's ChatGPT remains the most popular choice among newcomers, but for real, full-scale attacks, professionals increasingly turn to Google's Gemini. This suggests that Google's models apparently provide more flexible capabilities or less stringent restrictions in the context of complex tasks.
Methods for bypassing built-in model restrictions are actively discussed, and some enthusiasts are already attempting to create full-fledged malicious applications using publicly available AI services.
A New Era of Ransomware
The ransomware landscape has also undergone significant changes. Over a year and a half, more than 50 new groups have emerged, and the number of related incidents has exceeded 9,300. The key trend is the abandonment of data encryption as the primary method of pressure. Instead, attackers have shifted to a tactic of pure blackmail with the threat of publishing stolen information. This reduces the technical complexity of the attack but makes life many times harder for security teams.
The stolen data market continues to function: despite the closure of six major shadow marketplaces, five new ones have emerged in their place. On active card shops, more than 16.5 million compromised payment cards are for sale. Notably, even OpenAI has been forced to slow down the development of its advanced systems after a series of incidents, confirming the seriousness of cyber threats to the AI industry itself.
My comment: We are witnessing the convergence of two trends—the democratization of hacking tools and the growing complexity of attacks. The use of AI allows attackers to scale their efforts and adapt in real time, making traditional defense methods increasingly ineffective. The security industry will have to move from a reactive model to a proactive one, using the same AI technologies for predictive threat analytics.