CrowdStrike Counter Adversary Operations identified four previously undocumented cyberattack campaigns utilizing autonomous AI agents. These incidents demonstrate a shift where attackers delegate decision-making processes to AI systems to conduct large-scale operations.
In Argentina, an attacker compromised at least 15 medical facilities using a dual-agent architecture based on PentestGPT. The system utilized a supervisor agent to plan tasks and an executor agent to perform them, resulting in the theft of credentials and medical records.
A Chinese-speaking actor targeted government entities and private companies across Brazil and the US, including firms in the biotech and retail sectors. The attacker accessed Claude models via a proxy to deploy systems with varying levels of AI autonomy.
An Indonesian hacker compromised government domains and private firms using publicly available GitHub tools. Despite limited security expertise, the attacker used AI to bypass defensive barriers, requiring only occasional manual intervention.
A Russian-speaking group breached a US CRM provider and various SaaS and data center companies in Russia, Kyrgyzstan, and the US. They implemented a seven-stage workflow that grants AI agents near-total autonomy in selecting hacking methods.
CrowdStrike warns that AI agents are commoditizing offensive cyber operations. The firm expects that within 12 months, low-skilled attackers will leverage these tools to conduct campaigns previously limited to state-sponsored or organized groups.
