OpenAI says prompt injection may never be ‘solved’ for browser agents like Atlas

OpenAI details rising prompt-injection risks against its ChatGPT Atlas browser agent and describes a recent security update featuring an adversarially trained model and an automated attacker that iterates with counterfactual rollouts to discover multi-step exploit paths. A hypothetical scenario shows a malicious email prompting the agent to send an unintended resignation letter, illustrating how agent-enabled workflows shift traditional risk. The report aligns with UK NCSC guidance that prompt injection may remain a persistent challenge and highlights OpenAI’s broader preparedness efforts to study and limit emerging AI-related cybersecurity risks.