Governments and industry have a limited opportunity to strengthen cyber defenses before advanced AI attack capabilities become widely available, OpenAI said in a policy blueprint published on Oct. 21.
The AI developer’s publication, “Democratic governance of frontier AI, part II: Empowering cyber defenders,” calls for stronger safeguards alongside wider access to defensive tools. OpenAI said its experience with security incidents and models behaving outside intended boundaries informed the proposals.
OpenAI said AI is changing how quickly cyber operations can run, how widely they can reach and what they cost. The company said the technology can help defenders find and repair software weaknesses, identify intrusions and respond to attacks.
Those same capabilities can also help malicious actors, the company said. It said AI systems create further risks when they act beyond the boundaries set for them, making safeguards and accountability necessary as defensive use expands.
The blueprint offers proposals for governments and industry rather than findings from a study covering a specified period. OpenAI supplied no numerical estimates of the opportunity’s duration, the scale of the risks or the benefits of its proposals.
OpenAI called for clearer responsibilities for protecting the most advanced AI models and agents, systems that can carry out tasks. It also proposed requirements to detect dangerous behavior and preserve dependable controls for shutting down operations within each organization’s control.
The company said defenders need access to advanced tools, training and technical support. It coupled that proposal with calls to fix longstanding cybersecurity weaknesses, protect essential services and replace outdated systems.
OpenAI also proposed stronger government testing of AI systems and independent technical assessments. It called for consistent requirements to report incidents, alongside better coordination when organizations respond to attacks and recover from them.
The proposals build on OpenAI’s earlier cybersecurity action plan and blueprint for the safety of advanced AI, the company said. It attributed the need for stronger industry safeguards partly to incidents involving its own models.
OpenAI said AI developers bear responsibility for addressing risks created by their systems and helping improve society’s defenses. The company said governments also need the technical expertise, resources and authority to protect the public as AI capabilities advance.
OpenAI called on policymakers to establish safeguards now and strengthen requirements as the technology changes. Its proposed approach also calls for learning from experience, with government oversight and industry responsibilities adapting as more capable systems become available.
