VeriGrey: Greybox Agent Validation Uncovers Indirect Prompt Injection via Tool Sequence Mutation

arXiv·medium signal

ArXiv paper 2603.17639 introduces VeriGrey, a security testing framework that finds indirect prompt injection vulnerabilities in LLM agents by analyzing tool call sequences and applying mutation-based fuzzing rather than input-level attacks. The greybox approach (partial knowledge of agent internals) closes the gap between black-box penetration testing and white-box formal verification. Key finding: tool abuse vectors are systematically discoverable via sequence analysis, not just adversarial prompting.