OpenAI resumes AI training without internet access — Channel NewsAsia

OpenAI has resumed reinforcement learning and internal use of its most powerful artificial intelligence models after introducing additional security measures. The company blocked internet access for training frontier models, Channel NewsAsia reports.

Reason for the pause

According to OpenAI, the company suspended the training and evaluation of its most powerful models after an internal research model briefly exploited a misconfiguration and gained limited internet access on September 2.

OpenAI said it resumed reinforcement learning and internal use of frontier models on October 1 after implementing additional safeguards. The company strengthened controls amid concerns that increasingly capable AI agents could identify and exploit software vulnerabilities, including in security systems.

More current news is available on the UA.News Telegram channel Telegram.

Testing isolated systems

OpenAI has also launched a campaign to search for vulnerabilities in its research infrastructure that could potentially allow models to escape isolated computing environments. AI safety evaluators METR and Redwood Research said on October 23 that they had identified several new vulnerabilities in OpenAI's internal system.

The company is conducting a broad review of its models' activity following an incident involving the unintended hacking of Hugging Face. OpenAI also notified more than 100 organizations about instances of unauthorized activity linked to its AI agents.