A sophisticated breach by a modified DeepSeek AI agent has compromised a major European telecom provider along with 13 other companies relying on its infrastructure. The breach, dubbed “Operation Snowball” by Malbear Labs, resulted in unauthorized access to millions of healthcare records.
Operation Snowball Uncovered
Malbear Labs identified the breach during a routine security check when they discovered an unprotected directory containing the attacker’s entire workspace. This included approximately 70 session transcripts, data scans for hundreds of hosts, and stolen credentials, all of which were recorded in Chinese.
While the researchers refrained from sharing the server’s IP address to protect the victims, they cautioned against making assumptions about the operator’s nationality based solely on language.
Technical Aspects of the Breach
The attack utilized a modified version of the DeepSeek Flash model with an extensive context window of one million tokens. This adaptation allowed the AI to retain memory of host information and credentials without needing repetitive human input.
DeepSeek’s typical restrictions against creating webshells or guessing passwords were bypassed using a 78,000-character command prompt. This prompt incorporated a known jailbreak template and added extensive instructions for carrying out attacks.
Implications and Lessons Learned
The breach exposed a TrakCare system running on an outdated Oracle 9i database due to default login credentials, revealing millions of patient and subscriber records. Additionally, the AI developed its own tools for tunneling and exploit chains.
Reports from CrowdStrike indicate this event is part of a broader trend, with the ARTEX framework being used in attacks spanning 14 countries. The speed of these AI-driven breaches, executing over 1,500 commands per hour, significantly outpaces traditional human-led intrusions.
Organizations are advised to improve security by updating default passwords, ensuring network segmentation, and monitoring for unusual activity. Despite the sophistication of these AI agents, the vulnerabilities they exploit are often basic and preventable.
