Prompt Injection Is a Permanent Feature of LLMs, Says Security Researcher

Jason Haddix argues that prompt injection cannot be eliminated under current transformer architectures, as instructions and data are indistinguishable to the model. He notes that even frontier models like GPT-5 can only mitigate, not solve, the problem, with industry leaders aiming for 98% reduction. The issue extends to agentic tools ingesting untrusted text, requiring layered defenses.