Phalanx vs the World

This platform provides a unified arena for comparing instruction-control systems. Key features include: • Two testing lanes: one unprotected, one with Phalanx • Challenges for direct jailbreaks and indirect prompt injection • Support for mixed multi-turn attacks • Public scoring protocol for attack verdicts • Secure testing of sandboxed mock actions It allows users to test the robustness of language model safeguards against various adversarial techniques. The arena tracks and displays public records, including labels, hashes, and verdicts for each attempt. Users can submit prompts and attach text files to explore the boundaries of instruction control. The platform aims to demonstrate the effectiveness of protection mechanisms by comparing outcomes between the naked and protected lanes. Designed for security researchers, developers, and those interested in evaluating language model integrity. It provides a transparent environment to assess and improve the resilience of text-based systems against manipulative inputs, ensuring secure and reliable operational contexts. The challenge invites participants to 'break Phalanx' and earn a receipt for successful attempts.