Pacing model development in an era of cyber-critical capabilities

A major AI developer temporarily slowed its model development pace after discovering that an upcoming model may have reached a critical level of cybersecurity capability, prompting concerns about risks from increasingly powerful AI systems. The company responded by strengthening safeguards across three areas: monitoring systems to detect concerning behavior, alignment work to ensure systems behave as intended, and security measures to limit what AI systems can access. These changes include isolating research environments where frontier models are trained, implementing network controls, and expanding monitoring to catch unauthorized activities like data theft or attempts to defeat safeguards. The developer framed these actions as necessary because the risks associated with developing and testing more capable AI systems have grown faster than existing oversight standards.