OpenAI tightened safeguards for GPT-6.2 Astra and GPT-6.2 Sol, reducing the success rate of extracting protected internal reasoning from 54 percent to 0.2 percent in automated tests. The defense combines targeted training, message filters, API blocks, and weekly bans on thousands of abusive accounts