Ain't PII if it ain't SSN

OpenAI's GPT-6 Luna was tested with 9 adversarial PII traps. It never touched US SSNs, but took passport information in 6 out of 9 attempts, especially under Claude-Code and Codex-config runs. Aadhaar details were taken in 4 out of 9 attempts. Credit card information was never taken, with 4 refusals for pre-authorization variants. The model appears to understand PII protection, but applies it differently based on nationality, as passport data was taken without hesitation, unlike US SSN or credit card information.

Unlike other models that fail universally, GPT-6 Luna uniquely applies PII protection by nationality, accepting passport data while refusing US SSNs and credit cards.