prompt injection

OpenAI's GPT-6 Astra, released on September 3, marks a measurable improvement in factual accuracy and resistance to direct prompt manipulation compared with its predecessor, GPT-5.6 Sol. But independent security testing reveals the model can still be compromised through hidden instructions…