Prompt Injection Is a Permanent AI Architecture Flaw, Says Security Expert Jason Haddix

Security researcher Jason Haddix, speaking on the Secure Disclosure podcast roughly two months ago, argued that prompt injection attacks cannot be fully eliminated as long as large language models rely on current transformer-based architectures. Because these models treat instructions and data as identical text, there is no structural boundary to prevent malicious input from being interpreted as a command. Haddix noted that even leading AI figures like Dario Amodei and Sam Altman speak only of reaching around 98% mitigation, not complete prevention — a ceiling he compared to the imperfect but manageable state of web security today. He observed that while early jailbreak tricks have largely stopped working on frontier models, attackers now combine multiple techniques to bypass layered defenses including safety training and classifiers. His core advice for developers building agentic AI systems is to treat prompt injection resistance as an ongoing security discipline rather than a problem awaiting a permanent fix.