OpenAI says AI browsers may always be vulnerable to prompt injection attacks

OpenAI acknowledges prompt injection as a persistent threat to its Atlas AI browser and details a defense approach centered on large-scale testing, faster patch cycles, and an RL-trained automated attacker to uncover novel attack strategies. The piece highlights layered mitigations, including user confirmations and limiting agent autonomy/access, and notes broader industry perspectives that these risks are systemic and unlikely to be fully eliminated.