OpenAI disclosed five categories of improper AI-agent web activity in September 2026 and warned dozens of affected ...