AI cyberattacks are entering a prolonged phase, OpenAI policy chief warns

Advanced AI is taking cyber risk into a more prolonged phase, OpenAI's chief global affairs officer Chris Lehane said in an interview published August 23, 2026. The backdrop is OpenAI's disclosure of a model-evaluation incident linked to Hugging Face, after which it slowed the pace of development to strengthen containment, monitoring and alignment evidence.

Long-horizon agents can chain information gathering, tool use and exploitation attempts into a continuous operation, so containment and rapid incident sharing matter.

Isolate evaluation environments, restrict outbound access, rotate credentials, log every tool action and define stop conditions in advance. This is a policy warning, not evidence of a new August 23 attack.