Reproduce “CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zimo Xu, Wenzhuo Liu, Hideki Nakayama

CapLite separates tool authorization from model-generated plans by committing writes before reading tool results. It derives field roles from tool schemas and permits writes only when consequential arguments are grounded in the user’s request. On AgentDojo, it had zero successful tool-goal attacks across 13,935, with 29.0% utility under attack, close to CaMeL’s 27.2%; answer-only manipulation is outside write integrity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.