Safety filtering does not produce an error status. When the prompt itself is blocked the response has an empty or absent candidates array and a promptFeedback object carrying blockReason. When the output is blocked mid-generation you get a candidate whose finishReason is SAFETY, or another non-STOP value such as RECITATION or MAX_TOKENS, and whose content may have no usable parts.
Helper accessors that return the first text part throw or return nothing in both cases, which surfaces as a null-reference crash rather than as a content policy event.
Always read finishReason before reading text, and check promptFeedback when candidates is empty. Log blockReason and the safety ratings so the difference between a filtered request and a genuine failure is visible in production. Safety thresholds are configurable per category through safetySettings on the request, but the None and Off thresholds are not available for every category or every deployment, so verify the setting took effect instead of assuming it did.