MetaPI Turns Agent Benchmarks into Prompt Injection Testbeds
AI-generated summary A new paper presents MetaPI, a framework that converts existing agent benchmarks into prompt injection test environments by finding injection points, generating attack goals, optimizing adaptive payloads, and verifying attacks in real execution environments. The authors evaluated MetaPI on AppWorld, VitaBench, and Terminal-Bench, noting that closed frontier models such as gpt-5.6-luna and claude-sonnet-5 remained the most robust targets while still recording attack success rates above 20 percent.