OpenAI Unveils GPT-Red AI Model That Automatically Finds Prompt Injection Vulnerabilities

OpenAI unveiled GPT-Red, an internal automated red‑teaming model that generates adversarial prompts to identify prompt‑injection weaknesses in agentic AI systems; in controlled tests it successfully induced data‑exfiltration scenarios and manipulated an AI‑powered vending agent, and OpenAI is using those attacks as adversarial training data while keeping GPT‑Red separate from deployed products.