OpenAI acknowledges that prompt injection attacks on AI agents like its ChatGPT Atlas browser are a persistent, unsolvable threat similar to web scams. The company is combatting this with an LLM-based automated attacker bot, trained via reinforcement learning, to proactively discover novel attack strategies in simulation. While hardening defenses with rapid testing cycles, OpenAI and experts note AI browsers high access to sensitive data creates significant security trade-offs.