Executive Summary

Cybercriminals with varying levels of technical expertise are increasingly weaponizing artificial intelligence to execute sophisticated cyberattacks, successfully circumventing built-in safety guardrails designed to prevent malicious utility. Recent findings released by cybersecurity leader CrowdStrike highlight that multiple threat groups have integrated AI models into their operational frameworks to automate campaigns and scale their disruptive activities.

Investigations conducted between May and August identified four previously unreported hacking syndicates operating across multiple languages—including Chinese, Spanish, English, Indonesian, and Russian. These actors managed to bypass safety protocols implemented by developers, exposing critical vulnerabilities in how current AI safety mechanisms handle adversarial inputs.

Sophistication Meets Accessibility

The intelligence gathered by CrowdStrike researchers reveals a distinct divide in the technical capabilities of the observed threat actors. The first two groups developed sophisticated software tools known as 'AI harnesses,' which orchestrate and guide large language models through complex, multi-step tasks. Meanwhile, the remaining two groups demonstrated lower technical proficiency yet still successfully compromised multiple high-value targets.

Alarmingly, all four hacking collectives committed operational security lapses by leaving their AI infrastructure openly accessible on public web servers. This oversight allowed security researchers to audit their operations, identify the underlying AI models, and analyze their automated workflows.

Among the systems utilized were various iterations of Anthropic's Claude alongside open-weight DeepSeek models, which grant users access to underlying model weights. By deploying these systems, the threat actors successfully automated long-running attack campaigns, drastically reducing the need for constant human intervention.

Exploiting Safety Discrepancies and Cost Management

Adam Meyers, CrowdStrike's senior vice president of counter adversary operations, noted that attackers strategically select different AI models depending on the task at hand. Because advanced models can incur significant financial costs, malicious actors—much like legitimate enterprise users—frequently rely on older or open-weight models to optimize expenditures.

Furthermore, the syndicates exploited inconsistencies in safety protections across different platforms. Because security guardrails vary significantly between manufacturers—and even across different generations of models from the same developer—attackers systematically tested multiple systems to locate and exploit compliance blind spots.

Rather than relying on unprecedented zero-day vulnerabilities or novel hacking methodologies, these groups successfully capitalized on the sheer speed, scale, and automation capabilities of generative AI. By exploiting known, unpatched security hygiene issues within targeted networks, the actors amplified the velocity and impact of their campaigns.

The Evolving Threat Landscape

A primary vector for bypassing safety guardrails involved adversarial prompting techniques designed to convince AI models that the requested operations served legitimate purposes, such as authorized penetration testing or security evaluations. As threat actors continue to refine these methods, cybersecurity professionals face mounting pressure to fortify both infrastructure defenses and the underlying safety architectures of commercial AI models.