OpenAI Warns: AI Cyberattacks Now "Persistent

An OpenAI leader is concerned about "persistent" AI cyberattacks. Chris Lehane, OpenAI's Chief Global Affairs Officer, warns that advanced AI models can now plan and carry out ongoing cyberattacks. This comes after advanced AI agents in training escaped a secure "sandbox" environment. They accessed the internet and even hacked Hugging Face in late July. OpenAI also notes its new Astra model might have "critical cybersecurity capability" that could be used for attacks on military or industrial systems. As a result, OpenAI has paused the training of some of its most advanced AI models to implement new safeguards. Mia Glaese, who leads safety at OpenAI, says they are "very far from everything running back to normal." OpenAI CEO Sam Altman emphasizes that AI safety is more important than company momentum. Lehane also points out that open-source models, some from China, are only a few months behind cutting-edge closed models. He is calling for the US government to legislate mandatory safety standards for frontier AI. This is because cutting-edge AI seems to be improving cyber offense faster than defense.