WDTA’s new standard is part of its broader AI STR (Safety, Trust, Responsibility) certification suite. It provides a systematic approach for testing, validating, and certifying AI agent behavior.

Key features include:

• Two-layer testing framework:
  The standard combines Agent System Security Testing and Lifecycle Security Testing — evaluating not just the architecture (interface, model, tools, memory, and RAG systems), but also every phase from development to deployment and maintenance.

• Comprehensive threat modeling:
  It identifies and classifies risks unique to agentic AI, such as prompt injection, memory poisoning, RAG data leakage, model stealing, and backdoor attacks — going far beyond traditional software vulnerabilities.

• Real-world attack simulations:
  Includes adversarial test scenarios like jailbreaking, knowledge base poisoning, malicious tool invocation, and unauthorized memory access, ensuring agents are evaluated under realistic, high-stakes conditions.

• Quantifiable risk evaluation:
  Defines clear assessment criteria and metrics (e.g., failure rates under attack, response compliance, leakage probability) to measure how securely an agent behaves in dynamic, unpredictable environments.

• Focus on single-agent autonomy:
  This is the first known standard globally to focus on single-agent systems — where one AI agent acts independently using memory, tools, and models — a configuration increasingly adopted across industries.