Developer builds free tool to test LLM chatbot system prompts against five injection attacks

A developer at AI company Framz has released a free, no-signup tool that tests the resilience of large language model system prompts against five categories of prompt injection attacks. The tool covers instruction overrides, prompt extraction, delimiter escapes, role-play jailbreaks, and indirect injection via retrieved documents or tool calls. Indirect injection was highlighted as particularly dangerous, as models cannot distinguish legitimate instructions from malicious ones hidden inside fetched documents or RAG content. The tool runs on Framz's own infrastructure, meaning users' system prompts are not sent to third-party models for analysis. The developer cautions that the five attack classes represent a baseline check rather than a comprehensive security guarantee, as prompt injection remains an unsolved problem.