OpenAI says prompt injection in browser agents is “unfixable.” Here’s what actually helps.
OpenAI has stated that prompt injection in browser agents is an unfixable structural vulnerability at the model level. However, a new architectural solution called Arc Gate has demonstrated significant success in mitigating these attacks. Arc Gate reportedly achieved 100% effectiveness on the AgentDojo benchmark and 99% on InjecAgent, with independent verification showing a 25/25 success rate compared to the same model without the protection. AI
IMPACT Provides a potential architectural fix for prompt injection vulnerabilities in AI agents, enhancing their security and reliability.