Summary
This report is a technical document that outlines the current high-probability threats affecting enterprise use of Large Language Models (LLMs) and provides recommended countermeasures. It addresses threats targeting both self-hosted/internal LLM applications and LLM functionality embedded into SaaS products. The report identifies several key threat patterns, including indirect prompt injection via enterprise content, over-permissioned connectors, tool/agent abuse, knowledge base poisoning, operational data leakage, and supply chain risks. It emphasizes that the most significant risks arise from LLMs acting as interfaces to enterprise data. The document includes trending threat scenarios, a mapping approach using MITRE ATLAS and ATT&CK frameworks, and a control baseline for prevention, detection, and response strategies. Additionally, it provides a monitoring and incident response appendix, highlighting the limited attribution in public reporting for LLM-specific attacks and the importance of addressing standard intrusion techniques combined with misconfigurations.
