According to the company, the agents made decisions and executed commands rather than only producing tools or other artifacts. The operators exposed offensive agent frameworks through open directories.
One Spanish- and English-speaking operator allegedly compromised at least 15 healthcare entities in Argentina. CrowdStrike said credentials and medical records were exfiltrated from at least one victim. The operator used a dual-agent penetration-testing harness: a Claude Code “Supervisor” selected the next state-machine task, while an “Executor” performed it.
CrowdStrike described the framework as derived from the open-source PentestGPT project and modified with MCP servers, Python exploitation scripts, and skills. The report does not establish that these four actors were coordinated, that they explain a wider set of recent security incidents, or that AI operated without human direction.