Prompt injection isn't the bug — the framework holding the keys is

The article argues that prompt injection is not a fixable bug but an inherent property of LLM-based systems. The real vulnerability lies in AI agent frameworks that give models tool access and autonomy without meaningful separation between trusted and untrusted input. The author suggests the industry should stop chasing prompt-injection patches and instead rethink how agent frameworks are designed, treating the model as an untrusted reasoning engine rather than a trusted executor.