AgentFix — AI Agent Prompt Security

AgentFix is a full-stack AI agent prompt security platform built with IBM Bob IDE and Vertex AI Gemini 2.5 Flash. It solves a critical gap in AI development — developers have no way to test their agent prompts for security vulnerabilities before production deployment. The platform provides a complete 5-step pipeline: ANALYZE — Scans all 8 sections of your prompt pack across 40+ issue types. Generates a reliability score across 13 security dimensions with severity ratings and actionable fix recommendations. ATTACK MODE — A two-lane red team engine. Lane 1 uses Gemini 2.5 to generate intelligent context-aware attack scenarios mapped to 3 simulation profiles and 14 named test IDs. Lane 2 pulls from 600+ real adversarial payloads from the HuggingFace prompt injection dataset, scored by complexity and selected with hard-payload bias. Supports multi-turn attacks up to 3 conversation turns. Live Agent Probe mode fires attacks at a real Vertex AI agent and captures actual responses as evidence. SELF-IMPROVE — Iterative prompt hardening engine that reads every failed attack, understands the root cause, and automatically rewrites the vulnerable sections. Runs up to 5 improvement iterations until a reliability threshold is met. VERIFY IMPROVEMENT — Reruns the exact same attacks against the improved prompt. Shows before and after reliability scores, failure count reduction, and vulnerability reduction percentage. REPORT — One-click developer report exported as PDF covering all issues, attack results, improvements made, and remaining risks. Built entirely with IBM Bob IDE as the primary development partner. Full session documentation available in the repository. Powered by Vertex AI Gemini 2.5 Flash via streamGenerateContent SSE with built-in observability, retry logic, and Express Mode support.