What happened
RAND's Center on AI, Security, and Technology presents a structured, vulnerability-centric framework for identifying and characterizing security weaknesses in generative AI systems, moving beyond conventional attack taxonomies. The authors systematically decompose AI architectures—from training data and tokenization through transformer layers and deployment interfaces—to map where and how vulnerabilities arise, identifying '31 distinct classes of AI vulnerabilities,' most of which differ fundamentally from traditional software flaws because they emerge from probabilistic learning dynamics, data composition, and optimization trade-offs rather than deterministic code errors. The report finds that many of these vulnerabilities are only partially patchable, requiring architectural safeguards, provenance validation, and continuous monitoring rather than conventional software updates, and it recommends integrating AI-specific vulnerabilities into global standards for systematic risk management. Authored by Elie Alhajjar, Sasha Romanosky, Kyle A. Kilian, and Joe Uchill, and published July 30, 2026.
Why it matters
CISOs and security architects need a structural (not just adversarial-technique-based) taxonomy to scope AI red-teaming, vulnerability disclosure, and patch-management programs, since the report's core finding — that most AI flaws can't simply be patched like software bugs — has direct implications for how enterprises budget for continuous monitoring versus one-time remediation.
Action needed
Map current AI red-teaming and vulnerability management programs against the report's 31-class taxonomy to identify structural coverage gaps that traditional software vulnerability scanning would miss.