Semantics-based Prompt Injection Prevention Tool

Semantics-based Prompt Injection Prevention Tool\n\nObjective: help prevent prompt injections by combining semantic similarity checks with a probability-based risk rating for each prompt.\n\nBackground: A previous side project was compromised via clever prompt injections, burning through API credits. This tool is built to help others avoid the same fate.\n\nHow it works: For every candidate prompt, compare its semantics to known injection patterns and related risk factors. Compute a threat score (0-1) using a probability-based rating. Return a concise report including: threat_label (e.g., HIGH/MEDIUM/LOW), score, observed injection vectors, recommended mitigations, and any edge cases.\n\nCurrent status: Not perfect yet; observed detection effectiveness around 97%, with plan to reach 99.7% using an LLM-in-the-loop system.\n\nWhat we ask you to do: Test the tool by running a diverse set of prompts (including adversarial, ambiguous, and benign prompts). Provide feedback, share edge cases, and try to break the system to help improve robustness.