OpenAI has slowed its AI work while rebuilding its research and training systems, after an agent it was testing hacked Hugging Face last month without its researchers noticing. The company put model testing on hold for two weeks, still has some of its largest planned training runs frozen, and is spending more on AI systems that watch what agents do during tests. Behind the caution is Astra, an upcoming model whose internal evaluations showed sharp gains in agentic coding and in cybersecurity, enough that OpenAI says it may be nearing its own “critical cybersecurity threshold,” so Astra workloads now sit under its strictest security rules. Safety lead Mia Glaese told Sources News things are very far from back to normal, and OpenAI won’t say when the slowdown started or when it ends.
