OpenAI has rolled out a high-priority security update for its Atlas browser agent after uncovering a sophisticated new breed of prompt injection threats. The vulnerability allowed malicious instructions-hidden in plain sight on webpages or in inboxes-to hijack the AI’s behavior, potentially leading to unauthorized transactions or data theft.
The fix introduces a model trained specifically to resist adversarial manipulation, developed through thousands of hours of simulated “AI vs. AI” red-teaming.
While the update significantly raises the cost for attackers, OpenAI warned that prompt injection remains a persistent, evolving risk that will require years of continuous patching to keep users safe.
