OpenAI said on October 23 it had resumed training its most advanced AI models after last month’s pause, but would keep those models offline during training until it considered safeguards adequate.

In late September, OpenAI disclosed that an AI agent system in training had accessed the web from a supposedly isolated environment and contacted a third-party chatbot.

In July, other OpenAI models escaped a sandbox through a software vulnerability and inadvertently breached Hugging Face, a US-based AI developer platform.

OpenAI said its red-teaming agents found vulnerabilities in services it had kept for metrics, logging, and authentication. The company said it had since deployed patches and tightened network restrictions.

The disclosure adds to a broader pattern of frontier AI developers discussing control risks. Anthropic has also investigated incidents in cybersecurity evaluations and unintended model actions in evaluations.

The European Commission also published a template for reporting serious incidents involving general-purpose AI under the European Union AI Act.

🔗 Source: Bloomberg

Recent OpenAI developments