Cybersecurity researchers have identified nearly two dozen security vulnerabilities across fifteen (15) machine learning (ML) open-source projects, posing significant risks to organizations. According to JFrog, these flaws affect both server- and client-side components, allowing attackers to compromise critical systems like ML model registries, databases, and pipelines. Notable vulnerabilities include directory traversal in Weave, tracked as CVE-2024-7340, enabling privilege escalation, a command injection vulnerability in Deep Lake, tracked as CVE-2024-6507, and a privilege mismanagement in Mage AI, tracked as CVE-2024-45187. Some issues, such as improper access control in ZenML, lack CVE identifiers but remain just as severe. Exploiting these flaws could enable ML model backdooring, data poisoning, and unauthorized system access. This disclosure follows JFrog’s earlier report on more than twenty (20) vulnerabilities in MLOps platforms and introducing Mantis, a defensive framework leveraging prompt injection to neutralize attacks on large language models (LLMs) with over 95% effectiveness. Mantis uses decoy services and dynamic prompt injections to autonomously disrupt or hack back attackers, highlighting the escalating cyber threats to ML systems. Administrators responsible for maintaining these projects should read the full report for more information and be on the lookout for patches.