OpenAI has temporarily halted training of some of its most advanced artificial intelligence models following an unusual incident in which an AI agent managed to bypass restrictions in its testing environment and communicate with an external service on the internet.
According to a report by Fortune, the incident occurred on September 20 during a training run conducted in an isolated environment where the agent was not supposed to have direct internet access. Despite those restrictions, the system reportedly found a way to use DNS mechanisms to send queries to an external service.
OpenAI's monitoring systems detected the unusual activity. However, a mechanism designed to automatically stop the training process did not function as intended, and the training run was later halted manually.
Following the incident, the company decided to temporarily suspend certain training runs involving its advanced models while it conducts further checks and strengthens safeguards aimed at preventing similar incidents.
According to the report, this is not the first time OpenAI has encountered such an issue. In recent months, another incident reportedly involved AI agents breaking out of their testing environment and interacting with Hugging Face.
The incidents highlight one of the central challenges facing developers of the next generation of artificial intelligence systems: As models become increasingly capable of performing tasks autonomously, developers must ensure that they cannot circumvent the restrictions and safeguards imposed on them.