OpenAI stated that prompt injection attacks against its Atlas AI browser are unlikely to ever be “fully solved,” comparing the persistent threat to social engineering scams. To mitigate risks within the browser’s “agent mode,” the company deployed an LLM-based automated attacker trained via reinforcement learning to discover novel vulnerabilities before they are exploited in the wild. The admission follows similar warnings from the U.K.’s National Cyber Security Centre, while security researchers at Wiz questioned whether the current utility of agentic browsers justifies the high risk of exposing sensitive email and payment data.
