What happened
On August 3, 2026, OWASP's GenAI Security Project published the OWASP Top 10 for LLM Applications 2026 (v1.0), a major revision of the widely-adopted LLM risk taxonomy. For the first time the ranking is informed by a corpus of 7,714 real-world LLM security incidents (6,639 classifiable) alongside the traditional community vote (weighted 75%/25%). Key ranking changes: Excessive Agency jumps from #6 to #3 (biggest mover, reflecting agentic/tool-use risk), Unbounded Consumption rises from #10 to #6 (reframed around cost-asymmetry/denial-of-wallet rather than just DoS), Misinformation moves from #9 to #7, Improper Output Handling falls from #5 to #10, and System Prompt Leakage is renamed/broadened to 'Hidden Context Exposure' (#7→#8). The guide explicitly cross-references NIST, MITRE ATLAS, CWE, and the OWASP Top 10 for Agentic Applications.
Why it matters
This is the reference taxonomy for LLM/GenAI application security used across the industry (vendors, CISOs, auditors, red teams) to prioritize controls and map assessments. The methodology change (incident-data-informed ranking) and the elevation of Excessive Agency/Unbounded Consumption directly reflect the shift toward agentic AI risk that dominated the reporting window, and will drive re-prioritization of AppSec/AI-security programs industry-wide.
Action needed
Map existing LLM/agentic AI security controls and threat models to the revised 2026 ranking; prioritize least-privilege/agency controls (ASI-aligned) and cost/consumption guardrails; update vendor questionnaires and red-team test plans to the new categories.