As OpenAI rolls out new defenses for its agentic browser, the company admits that prompt injection attacks may be an indefinite vulnerability for AI systems.
