OWASP Publishes 2026 LLM Risk List and Adds a Standard for Controlling AI Agents

OWASP has released its 2026 Top 10 for LLM Applications and added the Agent Control Standard to its GenAI Security Project. Together, the materials separate two security jobs that organizations increasingly need: prioritizing the risks around language-model applications and placing enforceable limits on agents while they run.

The Top 10 is OWASP’s flagship guidance for identifying and mitigating critical risks in applications powered by large language models. Its 2026 edition adds updated rankings and broader threat coverage, drawing on research from thousands of real-world AI security incidents and contributions from hundreds of AI security experts.

The Agent Control Standard, by contrast, is intended to extend the project’s existing work on agent risks, controls, identity, governance and testing toward practical runtime enforcement. The distinction is consequential: a risk list can help a team decide what to address, while runtime controls are meant to constrain an agent’s access, decisions and actions as it operates.

OWASP is also positioning the Top 10 as a bridge to established programs rather than a standalone AI-security checklist. The 2026 version maps its guidance to NIST, MITRE ATLAS, CWE and OWASP’s Top 10 for Agentic Applications. A new GenAI Security Industry Framework Crosswalk is designed to connect the project’s guidance with established security, risk and compliance frameworks.

The supporting resources

• An expanded AI Security Solutions Directory now offers views of Generative AI Security, Agentic Security, and AI and Agentic Red Teaming.
• The framework crosswalk is an open resource intended to help organizations relate OWASP’s GenAI guidance to their existing security, risk and compliance frameworks.

OWASP said the new Top 10 passed 10,000 downloads in its first 48 hours. The project also said it has surpassed 30,000 LinkedIn members, and that F5 and WitnessAI joined as Gold Sponsors while Evoke Security and Mondoo joined as Silver Sponsors.

The release offers a fuller set of materials for teams moving from risk identification to agent governance and runtime enforcement. But the announcement does not establish how broadly organizations will implement the controls, or how the donated standard will work across different agent systems; those questions will determine whether the framework becomes operational practice rather than reference guidance.