Generative AI systems are vulnerable to crafted prompts that bypass safeguards, leak data, or induce malicious actions. We curate a living corpus of real‑world injection strings, automate fuzz testing across major LLM APIs, and prototype semantic filters and policy‑based output rewriting. Deliverables include a public “Risk Atlas,” mitigation guidelines for Bangladeshi fintechs and government chatbots, and conference presentations on guardrail engineering.
