Devs hit Catch-22 testing safety filters

Safety testing shouldn't be a fireable offense for developers; current "refusal-by-default" filters create dangerous blind spots for critical crisis-routing apps. Cloud providers lack transparent "research mode" toggles for legitimate adversarial testing, while local models often refuse safety-related instructions due to hard-coded RLHF alignment. Azure OpenAI "Limited Access" remains a viable path for bypassing standard filters, and synthetic "jailbreak" datasets offer a safer alternative for early-stage testing as proactive whitelisting remains largely undocumented.