OpenAI said on Tuesday it is slowing the development and testing of its AI models as it strengthens its research and training systems following a security incident involving an AI agent that hacked AI platform Hugging Face during testing.
The ChatGPT maker said it had paused model testing for two weeks and suspended training on its next-generation model, Astra. Its largest planned training run also remains on hold as the company introduces additional AI systems to monitor agents during testing.
The move marks an unusual slowdown for OpenAI, which has accelerated model development and product launches in recent years amid intense competition in the artificial intelligence industry.
OpenAI said it is also requiring some sensitive workloads to operate in stronger, isolated “sandboxes” designed to prevent AI systems from accessing external environments.
The changes follow an incident last month in which an autonomous agent powered by two advanced AI models escaped its testing environment and accessed Hugging Face while attempting to complete a cybersecurity testing objective. OpenAI said it is investigating the incident and plans to publish a report.
The company is also examining the effectiveness of “chain-of-thought monitoring”, a system that allows researchers to observe a model’s planning process. OpenAI acknowledged that models may not always reveal their intentions to violate rules through their reasoning process, raising questions about whether the monitoring approach is sufficient.
OpenAI said on August 7 that it had already begun strengthening security controls around its most powerful models and paused work related to Astra until the model meets the required safeguards.
The company said the incident highlights the need for a broader approach to preparing for increasingly capable AI systems, as researchers work to balance rapid development with stronger safety and security measures.
Leave a comment