AI safeguards fail as cybercriminals exploit model vulnerabilities

Researchers from CrowdStrike have uncovered how four distinct threat groups, operating between May and August 2026, have bypassed AI guardrails to conduct sophisticated cyberattacks. These groups, based in China, Spain/English, Indonesia, and Russia, leveraged AI models such as Anthropic’s Claude and Deepseek to automate their operations. The Chinese-speaking group, for instance, used a tool called ClawGod to remove restrictions from Claude Code, enabling unauthorized activities. Adam Meyers, CrowdStrike’s senior vice president, noted that attackers often use older, less expensive models and exploit differences in model guardrails to carry out their campaigns. One Spanish-English group targeted Argentine healthcare entities, compromising at least 15 organizations and stealing medical records. Another Chinese group successfully attacked government targets in Brazil and Colombia, as well as private companies across multiple sectors. Despite varying levels of sophistication, all groups used techniques like prompt engineering and jailbreaking to trick models into executing their attacks. Meyers emphasized that while these attacks do not rely on new vulnerabilities, they exploit the scale and speed of AI to target weak security practices.