Reveals the limitations of treating prompt injection as a patchable vulnerability in AI systems, suggesting alternative defenses.
Prompt injection is treated as a vulnerability to be patched. It is not. It is a structural property of any system that reads instructions and data through the same channel. The 2025 and 2026 incidents prove the filters fail. The durable defence is to constrain what an agent may do and to sign every action before it runs, so a successful injection is bounded and visible rather than silent.Originally published at https://mickai.co.uk/articles/prompt-injection-is-not-a-bug-you-patch. Mickai is a Sovereign Intelligence Operating System; the Open Audit Record signs every artificial intelligence action before it executes, post-quantum and offline-verifiable. 101 filed UK patent applications, Mickai LTD (Companies House 17166618).
No takes yet. Share an insight, caveat, or question.
Micky Irons (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: