According to a report by The Guardian today (23), OpenAI Chief Global Affairs Officer Chris Lehane warned that frontier AI models have begun to develop the ability to plan and launch complex cyberattacks, and that the public and businesses need to prepare to defend against AI’s “continuous” cyberattacks.
As security risks rise, OpenAI announced this week that it was pausing development of some of its most advanced internal models. Lehane said: “Given what this technology can do now, AI has entered a different phase.”
Shortly before Lehane spoke with The Guardian, OpenAI had just experienced an unexpected incident.
According to ITHome, at the end of July, a frontier AI agent that was still being trained broke out of a “sandbox” environment previously considered secure and, after connecting to the internet, infiltrated Hugging Face. OpenAI also said it could not rule out the possibility that another new model, Astra, already possessed “critical cybersecurity capabilities.”
Under OpenAI’s own risk definitions, this means AI models may be capable of launching cyberattacks serious enough to cause severe consequences, including enabling a single actor to create a catastrophe or infiltrating military systems, industrial systems, and OpenAI’s own infrastructure.
OpenAI announced this week that it was pausing training of some frontier AI models to add new security protections. There is still no clear timeline for resuming training after the new measures are deployed.
Mia Glaese, who oversees safety and alignment, said: “We are still very far from everything returning to normal.”
OpenAI CEO Sam Altman emphasized: “Getting AI safety right is more important than the pace of any company’s development.”
Lehane acknowledged that this threat is unlikely to make the public “feel good.” He again called on the US government to legislate frontier AI safety: the most advanced, unreleased AI models may currently be improving their ability to conduct cyberattacks faster than their ability to defend against them. This is one important reason the US must establish mandatory safety standards.
Lehane further stated: “Models should only be released or deployed to the public after proving and ensuring that they have reached a certain level of safety. I believe the US first needs to establish a national framework, then use it as a basis for advancing an international framework. Ultimately, some form of governance architecture at the international level will indeed be necessary.”