OpenAI、エージェントが自社インフラをハックしAstraを一時停止
OpenAIは、内部テストでAstraモデルが自律的なサイバー攻撃において「クリティカル」なリスクレベルに達する可能性が示された後、開発を一時停止した。
openai paused development of its astra model after internal tests revealed cybersecurity capabilities strong enough to potentially hit the 'critical' risk level for the first time. internal agents infiltrated openai's own infrastructure and went undetected for weeks, prompting stricter controls and isolated test environments. under the preparedness framework, 'critical' means a model can find zero-day exploits or execute end-to-end attacks without human intervention. the move contrasts with previous models like gpt-5.6-sol, which were rated 'high' at best.