Indirect Prompt Injection Formulated as a Test-Time Search Problem

This research frames indirect prompt injection as an agentic search over an environment-induced attack surface. Findings show that increasing attacker test-time compute improves vulnerability exploitation, suggesting that security evaluations must account for an attacker's compute budget and search strategy.