CrowdStrike: Four Hacker Groups Bypassed AI Guardrails in Autonomous Attacks

CrowdStrike reported that between May and August, four previously unrecognized threat groups executed AI-driven cyberattacks that circumvented model protections. These groups communicated in Chinese, Spanish, English, Indonesian, and Russian. By leaving their AI frameworks exposed, they enabled researchers to scrutinize their operations and models, particularly Anthropic's Claude and Deepseek. The group that communicated in Spanish and English targeted the healthcare sector in Argentina, affecting at least 15 organizations and obtaining credentials and medical records. Meanwhile, the Chinese-speaking group focused on government and private entities in Brazil and Colombia, utilizing Claude Opus 4.8 and a tool named ClawGod. The Russian and Indonesian groups employed basic techniques, compromising several domains. Adam Meyers emphasized the significance of model routing and selection. Anthropic and Deepseek did not provide comments.