Threat Intelligence Sources for AI System Security

The article, written by Viacheslav Mosin from ITMO's AI Talent Hub and ITMO AI Security Lab, provides an overview of key sources for gathering information security threats related to AI systems. It mentions that Nvidia CEO Jensen Huang called the launch of ChatGPT the 'iPhone moment of AI', and notes that by 2025, 88% of companies reported using AI in at least one business function, according to McKinsey. The sources reviewed include MITRE ATLAS, OWASP Top 10 documents (for LLM, Agentic, Agentic Skills, MCP), Google Secure AI Framework (SAIF), and Sber's Threat Model for AI Cybersecurity. MITRE ATLAS is praised for its high update frequency, structured tactics and techniques, and machine-readable GitHub repository, making it suitable for building threat models. OWASP lists are popular and integrated into security tools but limited to ten risks each. Google SAIF provides a risk map with components, risks, and mitigation measures, but with fewer threats and high-level abstraction. Sber's model, updated in July 2026, describes 37 threats and 51 implementation methods, distinguishing between external and internal attackers. The article likely concludes that MITRE ATLAS is the best source for comprehensive threat modeling.