ATTACKERS have been hijacking exposed AI endpoints, particularly those related to self-hosted AI software, to conduct offensive operations without needing special authentication. Research by Zenity identified incidents where attackers exploited inference endpoints from Ollama and LiteLLM, specifically via unsecured ports. Three different operators utilized this method for various purposes, including penetration testing and web reverse-engineering.
Recommendations for organizations to protect themselves include ensuring model back ends are not exposed to the internet, implementing real authentication mechanisms, and monitoring traffic to AI infrastructure. The trend highlights the growing sophistication of attackers in leveraging AI technology.