OpenAI Halts Astra Training Over Cybersecurity Risk
The pause shows OpenAI acting on its own July proposal, Pacing the Frontier, which asked the U.S. government to develop slowdown technologies and international governance for AI development.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
OpenAI paused training of its most advanced models on August 17 after its new model, Astra, approached a Critical cybersecurity capability level under company safety standards. The company halted reinforcement learning for the latest product-bound model for two weeks and suspended the largest planned training run, instead confirming safety through small-scale training and evaluations.
The halt follows a security incident called the OpenAI-Hugging Face incident, though Astra was not directly involved. OpenAI has since tightened isolation of environments where AI-generated code runs and restricted internet connections, with the strictest rules applied to Astra and cyber-related models.
Monitoring is also being strengthened. In reinforcement learning and evaluations that use tools, unauthorized access, data theft, and circumvention of safety measures will be automatically detected, and work stops if a serious problem is suspected.
The move overlaps with Pacing the Frontier, published in July, which asked the U.S. government to develop technologies that slow development speed and to establish international governance. Whether the pause holds depends less on OpenAI's stance and more on whether global competitors, including China, make the same decision.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.