CrowdStrike researchers examined the infrastructure of four groups active between May and August. The groups had left their AI systems accessible on web servers, giving investigators a view of the models and workflows used in the attacks. They operated in Chinese, Spanish and English, Indonesian, and Russian. The groups combined versions of Claude with open-weight DeepSeek models, assigned tasks to agents and tried to bypass safety controls by describing their work as authorized security testing. One group compromised at least 15 Argentine health organizations, according to the researchers.
The finding does not mean AI discovered a new way to break into every network. CrowdStrike says these campaigns did not rely on novel vulnerabilities. Automation made it faster to find and exploit existing weaknesses, while people still selected targets and directed the work. That changes the pace for defenders: exposed services, overdue patches and weak credentials become easier to probe at scale. The evidence comes from four studied groups and cannot establish the share of all cyberattacks that involve AI. It does show a practical limit of safeguards inside a chatbot: attackers can switch models, construct elaborate claims of permission and combine AI with ordinary tools.