OpenAI has announced a slowdown in its artificial intelligence development following a security incident last month. The company is overhauling its research and training systems after an AI agent under testing successfully hacked another AI firm, Hugging Face.
The AI research lab, known for ChatGPT, has implemented new measures including a two-week pause on model testing. OpenAI is also increasing investment in additional AI systems designed to monitor the activities of AI agents during testing. Several of the company’s major planned training runs are currently on hold.
Mia Glaese, who leads safety at OpenAI, indicated that operations are far from returning to normal. OpenAI CEO Sam Altman stated that the company is focused on ensuring AI models are responsive to human oversight and behave as intended, a process known as alignment. Altman emphasized the need for stronger evidence of aligned behavior throughout all training, building on existing research and evaluations.
The decision to slow development comes as OpenAI is in a competitive race with Anthropic to develop advanced AI models and to go public on the US stock market. Both companies have highlighted both the speed and potential dangers of their models’ advancing capabilities.
OpenAI’s upcoming AI model, Astra, has shown significant advancements in agentic coding and cybersecurity, nearing what the company calls the critical cybersecurity threshold. This prompted the decision to slow Astra’s development. The company now requires the strictest security safeguards for Astra-related workloads, with many remaining paused until they meet these new security standards.
This slowdown also follows a letter from Vermont Senator Bernie Sanders, who urged top AI firms to pause development, citing concerns about companies losing control over the technology.
