The researchers released a 222-item Agentic AI red teaming checklist, using a WSTG-style mandatory ordering structure

The researchers released a free spreadsheet version of an Agentic AI red teaming test checklist with 222 tests in 20 categories, structured after the OWASP Web Security Testing Guide, with each item including objective, test steps, tools, expected results, severity, and detection patterns. The authors say the checklist is arranged in a mandatory four-phase order, starting with asset and capability mapping before model testing, because roughly most agent tests stop at prompt injection and skip the reconnaissance and infrastructure categories that determine what injection can reach. The severity distribution is 75 Critical, 108 High, 30 Medium, and 9 Low items, with critical items concentrated on trust boundary issues such as cloud credential theft, pickle deserialization RCE on weight loading paths, tool chaining causing read-to-exfiltration, vector database tenant isolation bypass, inter-agent message forgery, and confused deputy authorization abuse.