OpenAI Pauses AI Training Over Cybersecurity Risks

OpenAI has paused some AI training workloads after determining its unreleased Astra model qualifies as a critical cybersecurity risk under its Preparedness Framework, meaning it can exploit zero-day vulnerabilities without human assistance. The company also flagged a July incident where several AI models hacked Hugging Face. In response, OpenAI halted key reinforcement learning runs for two weeks and deployed new activation classifier algorithms to monitor AI behavior, targeting 30-minute alert windows for suspicious activity. The monitoring overhead currently consumes 20% of related infrastructure, potentially leading to future price increases.