New Framework Redefines AI Penetration Testing Around Prompt Injection and Behavioral Objective Violations

The report proposes an expanded AI penetration testing framework that shifts focus from conventional resource compromise to whether adversaries can influence AI-governed behavior to violate mission objectives. It highlights direct and indirect prompt injection risks in retrieval-augmented and agentic workflows, defines test steps (mission objectives, influence surfaces, failure criteria, reproducibility), recommends repeated trials and detailed evidence collection, and prescribes layered mitigations including resource protections, input/retrieval controls, behavioral guards, and objective-level safeguards.