OpenAI Finds Self-Replicating Prompt Injection Attacks

OpenAI disclosed Friday that its GPT models are vulnerable to self-replicating prompt injections, an AI worm attack discovered in June during red-teaming of GPT-5.6. No real-world incidents occurred. OpenAI is training future models with GPT-Red to recognize and resist these attacks, though the technique could backfire by making models stealthier.