New research shows that AI's extended reasoning creates a security vulnerability, with extremely high attack success rates across major models including GPT, Claude, and Gemini. (Decrypt)