CrowdStrike finds four hacker groups that attack with AI and Claude
All four groups tricked the AI models, usually with the incorrect claim that the work was an approved security test.
CrowdStrike studied attacks made between May and August by four groups that no one reported before. The groups work in Chinese, in Spanish and English, in Indonesian, and in Russian. All four left their AI systems open on web servers. This means that CrowdStrike could see them. They used mostly Claude of Anthropic and the open-weight Deepseek models, and mostly older versions.
How they work
The groups use some models, and they select each model for one task. Adam Meyers of CrowdStrike says new models cost too much, and are too much for many tasks. He also says the guardrails, the rules that stop misuse, are not equal, also in the models of one company.
How they trick models
Many groups put files in the work folder of the agent to show that tasks are approved. Some use usual jailbreak prompts that tell a story of approved work. The less skilled groups sometimes use rude prompts to push the model.
What is not known
The source gives no count of victims. Meyers says the attacks use no new flaws. They use speed and volume against basic security errors that a team could prevent.
Insight: Attackers select the model with the weakest guardrail
Meyers says guardrails are not equal, also in the models of one company. A related study of CKA-Agent, an attack agent, found a success rate of more than 95% on 4 commercial models, with groups of safe queries. Attackers select models like buyers. This means that one strong guardrail is not sufficient. Your defense must also work when the model does not refuse.
Sources
Posted
