AI Agent Verification Standards Are Coming — Get Ahead With OWASP and NIST

Workday's Agent Passport ties every AI agent attestation to a public standard — OWASP's LLM Top 10, NIST's AI Risk Management Framework, MITRE ATLAS — rather than an internal checklist. Whatever platform you build or deploy agents on, mapping your own agents against these same standards now is worth doing before it becomes a procurement requirement someone else hands you.

What Each Standard Actually Covers

• OWASP's LLM Top 10 catalogs the most common real-world vulnerability classes in LLM applications — prompt injection, insecure output handling, training data poisoning, excessive agency, and others — with concrete examples, not abstract principles
• NIST's AI Risk Management Framework provides a structured process for identifying, measuring, and managing AI risk across an organization, mapping cleanly to how most enterprise risk committees already think about governance
• MITRE ATLAS catalogs adversarial tactics and techniques specifically against AI systems, functioning as an AI-specific analog to the well-known MITRE ATT&CK framework for traditional cybersecurity

Why Mapping to External Standards Beats an Internal Checklist

An internal risk checklist reflects what your own team thought to check for — useful, but limited by what you already know to worry about. External standards like OWASP's LLM Top 10 are maintained by a broader security community actively tracking new attack patterns as they emerge, which means mapping against them surfaces risk categories an internal-only review is more likely to miss.

A Starting Point You Can Apply This Week

Take your highest-privilege AI agent — the one with the broadest data access or the most autonomous decision authority — and run it through OWASP's LLM Top 10 checklist item by item. Most teams doing this for the first time find at least one gap they hadn't previously considered, typically around excessive agency (an agent authorized to do more than its actual use case requires) or insecure output handling (agent outputs consumed downstream without validation).