OpenAI’s IH-Challenge Makes Prompt Injection Harder, Not Impossible
OpenAI’s IH-Challenge matters because it turns prompt injection defense from a loose prompting practice into a trainable model behavior: the model is taught to follow a ranked instruction hierarchy, with system prompts above developer and user requests, and tool outputs at the bottom. That does not eliminate jailbreaks or hostile inputs, but it materially improves…