OpenAI President Greg Brockman publicly admitted his company underestimated its models' cyber abilities after an agentic system breached Hugging Face infrastructure. His new essay details internal defenses, a personal remediation experiment, and ten urgent steps for organizations to adopt AI agents before the defender's narrow window closes. Action now could flip the advantage to security teams.