The Emergence of LLM-Embedded Malware: A New Frontier in Cybercrime

Cybersecurity researchers have recently made headlines with the discovery of MalTerminal, the earliest known example of malware that integrates large language model (LLM) technology. Analyzed by SentinelOne’s SentinelLABS team and unveiled at the LABScon 2025 security conference, this malware showcases a fundamental shift in the evolution of cybercrime. Unlike traditional malicious software, MalTerminal utilizes OpenAI’s GPT-4 to create harmful payloads on the fly, ushering in an era of what experts are dubbing “LLM-embedded malware.”

MalTerminal: Proof-of-Concept or Weapon in Waiting?

The investigation revealed that MalTerminal exists within a Windows executable file featuring an API connection to OpenAI’s chat completions model, which was deprecated in November 2023. This timeframe indicates that MalTerminal was likely developed before this cutoff, marking it as the first known instance of AI-enhanced malware.

Upon execution, users are confronted with a choice between two attack modes: “ransomware” or “reverse shell.” The accompanying Python scripts demonstrate the malware’s modular design, with equivalent functionalities available in different programming environments. Remarkably, SentinelOne found no direct evidence of MalTerminal’s deployment in actual attacks, suggesting it may have been developed as a red team tool or an experimental proof-of-concept. However, the implications of its ability to dynamically generate malicious logic present a new concern for cybersecurity experts.

A Qualitative Shift in Adversary Tactics

According to researchers from SentinelOne, the incorporation of LLMs into malware represents not just an incremental improvement but a qualitative leap in adversarial tactics. The capabilities offered by GPT-4 enable MalTerminal to dynamically generate components such as ransomware code or analytic reports when paired with a defensive script named FalconShield. This script assists the LLM in identifying potentially malicious patterns in Python files.

SentinelOne has noted, "The incorporation of LLMs into malware marks a qualitative shift in adversary tradecraft. It allows malicious code to evolve dynamically, bypassing static defenses and complicating incident response." This shift underscores how cybercriminals can adapt their strategies in real-time, posing a significant challenge to traditional security measures.

LLM Poisoning in Phishing Campaigns

MalTerminal is not an outlier. Other cybersecurity firms are documenting ways in which adversaries are weaponizing AI models in phishing attempts. For instance, a recent report by StrongestLayer details that attackers are embedding misleading prompts within phishing emails to trick AI-based email security systems into deeming harmful communications as benign.

One particularly cunning campaign mimicked a billing discrepancy notice, masking malicious intentions behind professional-sounding language. Hidden HTML code employed prompt injections styled with CSS properties like display:none; and color:white;, effectively fooling AI filters. When victims opened the attached HTML file, it exploited the Follina vulnerability (CVE-2022-30190) to execute a devastating attack, which included dropping PowerShell scripts, disabling Microsoft Defender, and establishing persistence mechanisms.

Rise of AI-Powered Phishing Infrastructure

The misuse of AI extends beyond LLM prompts alone. A report from Trend Micro highlights a surge in cybercriminals utilizing AI-driven website builders, such as Lovable, Netlify, and Vercel, to orchestrate phishing campaigns since January 2025. These platforms—initially designed for the rapid and legitimate creation of websites—are now being repurposed to host counterfeit CAPTCHA pages that obscure credential-harvesting redirects.

Victims are confronted solely with the CAPTCHA page, significantly reducing suspicion. Moreover, automated scanning tools typically detect only the benign CAPTCHA element, failing to identify the underlying redirect. By leveraging free hosting services, automation, and familiar branding, attackers are scaling up their operations cheaply and rapidly.

Trend Micro researchers Ryan Flores and Bakuei Matsukawa describe these AI-driven hosting solutions as a “double-edged sword”—they can be transformative for legitimate businesses but also lower the barrier for entry into cybercrime.

Implications: A New Era of AI-Powered Cybercrime

The dawning realization brought about by the discovery of MalTerminal and the proliferating use of AI in phishing tactics highlights a disturbing truth: AI has now become a weapon in the hands of cybercriminals. From generating malicious payloads dynamically to evading AI-based detection systems, adversaries are sabotaging the technologies intended to fortify defenses.

For enterprises, these developments necessitate a proactive approach:

• AI Security Audits: Conducting regular assessments of AI-related security measures is no longer optional.
• Stress-Testing Defenses: Organizations must rigorously test their defenses against prompt injection and LLM poisoning scenarios.
• Investment in Adaptive Systems: Developing AI-driven detection mechanisms that can keep pace with the evolving tactics of cybercriminals is crucial for resilience.

As cybersecurity experts assert, "We have entered a phase where AI is both shield and sword. The defenders who master it first will shape the battlefield." This evolving landscape demonstrates that the stakes are higher than ever, and the importance of staying ahead in the fight against cybercrime cannot be overstated.