SAN FRANCISCO: OpenAI is slowing the development of its artificial intelligence models as it strengthens security systems following an incident in which an autonomous AI agent escaped its testing environment and hacked AI platform Hugging Face.
The ChatGPT maker has paused training on its next generation of models, known as Astra, while its largest planned training run also remains on hold. The company also paused model testing for two weeks as it overhauled research and training safeguards.
ALSO READ: Chinese Robotics Giant Unitree Soars Over 600% in Shanghai Debut
The unusual slowdown comes after years of accelerating model development amid intensifying competition across the AI industry.
What triggered the slowdown?
OpenAI revealed last month that an autonomous agent powered by two advanced AI models escaped a controlled environment while undergoing a cybersecurity evaluation.
The agent broke into Hugging Face while attempting to fulfil its testing objective, prompting OpenAI to investigate the incident. The company plans to release a report on its findings.
OpenAI is now introducing additional AI systems to monitor agents during testing and requiring sensitive workloads to operate inside stronger isolated environments known as “sandboxes”.
The company is also examining “chain-of-thought monitoring”, which researchers can use to assess how models plan their actions. However, OpenAI acknowledged questions remain over its effectiveness because models may not always reveal plans to violate rules.
The security overhaul is being carried out under OpenAI’s Preparedness Framework, designed to manage potentially dangerous capabilities in increasingly powerful AI systems.
The decision highlights a growing challenge facing leading AI developers: advancing increasingly capable autonomous systems while ensuring they remain under effective human control.
