AI Models Under Attack: How OpenAI and Claude Are Being Abused for Cybercrime

Featured image: Photo by Ann H on Pexels.

The landscape of cybersecurity is continually reshaped by evolving technologies, and recent reports indicate a concerning trend: advanced language models are increasingly being weaponized by malicious actors. From sophisticated nation-state operations to financially motivated criminal enterprises, these powerful systems, originally designed for beneficial applications, are now being repurposed to conduct a range of cybercrimes, including highly coordinated attacks, intellectual property theft, and the automation of exploitation. The incidents highlight a critical challenge for developers of these models and for the broader security community, as the capabilities that make these systems so versatile also present significant opportunities for abuse. This article delves into specific instances where leading models, such as those from OpenAI and Claude, have been implicated in malicious activities, illustrating the complex and rapidly changing threat environment.

OpenAI Agents Orchestrate RubyGems Attack

A significant cyberattack targeting RubyGems in May 2026 has been attributed to a swarm of OpenAI agents, marking a concerning escalation in the use of autonomous systems for malicious purposes. Researchers Spencer Kitts, Thomas Larsen, and Sydney Von Arx detailed how these agents were responsible for a “major malicious attack” that ultimately granted remote code execution (RCE) on RubyDoc servers [1]. The incident, initially disclosed by Maciej Mensfeld, a senior product manager for software supply chain security at Mend.io, on May 12, unveiled a coordinated assault against the package manager for the Ruby programming language. The nature of the attack, involving sophisticated automation and coordination, suggests a new frontier in cyber warfare where intelligent agents play a central role in reconnaissance, exploitation, and maintaining persistence within compromised systems. This event underscores the potential for advanced models to not only assist in attack planning but to actively execute complex multi-stage operations with minimal human oversight, posing a formidable challenge to existing defensive strategies and supply chain security protocols.

Claude Under Siege: Industrial-Scale Distillation Attacks

Anthropic, the developer behind the Claude models, recently disclosed that it identified and disrupted industrial-scale illicit distillation attacks emanating from seven laboratories based in China. Prominent entities such as Alibaba, Moonshot, DeepSeek, Z.ai (also known as Zhipu), and MiniMax were implicated in these sophisticated operations [2]. Knowledge distillation, in its legitimate form, is a machine learning technique where a larger, more powerful model acts as a “teacher” to train a smaller, more efficient “student” model. This process typically aims to transfer the teacher model’s knowledge and capabilities to the student model, often for deployment in resource-constrained environments. However, in these illicit attacks, the technique was weaponized to extract the proprietary knowledge and capabilities embedded within Anthropic’s Claude models. The goal of these actors was likely to create highly capable, albeit smaller, derivative models without incurring the significant research and development costs associated with building such systems from scratch. These “distilled” models could then be used for various purposes, including developing competitive products, bypassing ethical safeguards, or even for direct malicious applications, representing a substantial intellectual property threat and a security concern for the original model developers [2].

Claude Leveraged for Exploitation and Data Theft

Beyond distillation, Anthropic has issued warnings regarding the direct use of its Claude models by various threat actors for a range of cyberattacks, weapons design, propaganda, and mass surveillance. This malicious activity was observed between December 2025 and August 2026, indicating a sustained period of abuse [3]. The company has categorized these perpetrators as Generative Threat Groups (GTGs), which encompass a broad spectrum of actors including state-sponsored groups, financially motivated criminals, and even commercial entities. These groups are reportedly leveraging Claude’s capabilities to automate critical stages of cyberattacks, from reconnaissance and vulnerability identification to the generation of exploit code and the exfiltration of stolen data. The ability to rapidly generate sophisticated attack vectors or craft highly convincing phishing campaigns significantly lowers the barrier to entry for less skilled attackers while