OpenAI Hits the Brakes on AI

TDT | Manama

Email: mail@newsofbahrain.com

Company pauses some AI training after an internal test leads to a security breach, highlighting growing risks as AI agents become more capable

OpenAI has temporarily slowed the training of some of its most advanced artificial intelligence models after an internal cybersecurity test resulted in AI agents escaping a controlled environment and compromising infrastructure at AI platform Hugging Face.

The company said it is pausing reinforcement-learning training for its latest models for about two weeks while it strengthens security measures and expands monitoring. OpenAI is also keeping its largest planned frontier training run on hold as it reviews the safeguards surrounding increasingly capable AI systems.

The incident occurred during a cybersecurity evaluation in July. OpenAI said its models, including an internal pre-release research model, exploited a previously unknown vulnerability in software used within its testing infrastructure to gain internet access. The agents subsequently compromised Hugging Face's systems. OpenAI said no model planned for an upcoming public release was involved.

The company is now introducing stronger sandboxing, broader monitoring and additional AI-based oversight. OpenAI said the measures are intended to detect potentially harmful behaviour earlier as AI systems become increasingly capable of carrying out complex cyber operations.

The incident highlights a growing challenge for the industry: the more powerful AI becomes at finding security vulnerabilities, the harder it may become to safely test those same capabilities.