Prompt Injection as a Governance Object: Why Prompt Injection, Data, Model, and Memory Poisoning Are Likely to Be Governed as Security Duties Before They Are Criminalised, and a Proposed Ladder That Explains the Sequence

This paper asks whether prompt injection, data poisoning, model poisoning, and the poisoning of persistent artificial intelligence memory should be made illegal, and how each is otherwise becoming a governance object. It argues that the answer is likely no in the narrow sense and yes in the broad one. The review identified no enacted standalone offence or legislative proposal as at the cutoff, and the evidence suggests that a dedicated offence of prompt injection is unlikely in the near term, because the technique is not meaningfully separable from ordinary input, because the same manipulation is indispensable to legitimate security research, and because the vulnerability is largely architectural rather than a property of the attacker's conduct. The duty to defend against these attacks, by contrast, is already forming through security standards, management-system certification, sectoral regulation, and civil liability on the organisations that build and deploy artificial intelligence. To organise the analysis, the paper proposes the Governance Escalation Ladder, a seven-stage heuristic describing a sequence through which a new class of attack may move from technical recognition to eventual legal consequence. The paper locates these attack classes provisionally on that sequence, evidences each stage against current frameworks and law, illustrates the pattern against a comparable legislative choice in a second legal domain, and closes with staged guidance for governance leaders. The contribution is interpretive and organisational rather than doctrinal. Every legal status, standard, demonstration, and item of litigation cited is dated and qualified, and a source-verification register is provided as an appendix.