OpenAI publishes cybersecurity blueprint for AI defenders
OpenAI has released a plan that suggests ways for governments and businesses to improve cybersecurity as artificial intelligence (AI) advances, according to a statement from the company.
The document builds on OpenAI's earlier plans for cybersecurity and AI safety. The company said its own experience with AI models that did not behave as intended and with security problems involving its models helped shape the proposals.
The plan covers three main areas. First, it calls for making advanced AI safer and more secure. This includes setting clearer responsibilities for protecting the most advanced AI models, finding dangerous behavior, and keeping controls that can shut these models down. Second, it focuses on improving cybersecurity by giving defenders advanced tools, training, and technical support. It also calls for fixing long-standing weaknesses and investing in essential services. Third, it suggests making public institutions better able to handle cyber threats. This includes improving government testing of AI and security systems, getting independent technical experts to assess them, requiring organizations to report incidents in consistent ways, and improving how institutions work together to respond to attacks and recover from them.
OpenAI said AI can help defenders find and fix security weaknesses, detect break-ins, and respond to attacks. But the company also warned that attackers can use these same abilities for harmful purposes. AI systems can also create new risks when they act outside the limits they were designed to follow.
The company described this period as a "defender's window." It said this is a limited opportunity to give defenders access to advanced AI before powerful tools for launching cyberattacks become widely available.
OpenAI said AI labs are responsible for dealing with the risks their systems create. It also urged governments to build the knowledge, resources, and legal authority they need to protect the public as AI becomes more capable.