OpenAI says attacks that trick AI agents into doing harmful actions will likely remain a long-term security problem for AI browsers. To fight this, it built an AI system that acts like an attacker to find and help fix these weaknesses before they’re abused.