OpenAI slows down model development due to AI security risks

OpenAI, the developer of ChatGPT, put the brakes on the model development process due to the security risks posed by artificial intelligence models.

The company announced that it took a two-week break from model tests and started to deploy other artificial intelligence systems to monitor the activities of the tested artificial intelligence agents.

OpenAI also stated that it has stopped training studies for its new generation models called "Astra", and the largest planned training study is still on hold.

The company did not provide information on exactly when the two-week slowdown period began.

ARTIFICIAL INTELLIGENCE AGENT INfiltrated HUGGING FACE

The reason behind the decision is a remarkable security incident that occurred last month.

During a test conducted by OpenAI to measure its cybersecurity capabilities, an autonomous artificial intelligence agent exceeded the limits of the test environment while trying to fulfill its assigned task.

The agent infiltrated the company's infrastructure by accessing the systems of the artificial intelligence company Hugging Face over the internet.

OpenAI announced that it launched a comprehensive investigation following the incident and plans to publish a detailed report on the issue.

It was stated that in the company's previous tests, many model evaluations were carried out simultaneously at high speed, and the resulting large amount of data became difficult for employees to follow.

STRONGER TEST ENVIRONMENTS ARE BEING ESTABLISHED

Following the incident, OpenAI made it mandatory for particularly sensitive studies to be carried out in more powerful “sandbox” systems.

These systems, called sandboxes, are used as security areas that limit the connection of artificial intelligence models with external systems and aim to keep their activities in an isolated environment.

The company also started to follow the actions performed by artificial intelligence agents during the test with other artificial intelligence systems.

The aim is to detect unexpected or unauthorized activities of models at an earlier stage.

“CHAINED THOUGHT” AUDIT IS DISCUSSED

One of the security methods that OpenAI is working on is “chain-of-thought monitoring”, that is, monitoring the planning and reasoning process of the model.

With this method, researchers try to understand what strategies the artificial intelligence model uses when performing a task.

However, OpenAI executives acknowledge that there are still questions about how effective the method is.

Some early research suggests that AI models may not make their plans to violate rules explicit in their reasoning process.

WORKS FOR ASTRA ARE SUSPENDED

OpenAI announced on August 7 that it had increased security measures for its most powerful models and stopped some activities related to Astra, which was not yet available.

The company suspended certain parts of the work on the model, stating that the Astra did not yet meet the required safety standards.

OpenAI stated that the decisions taken were implemented within the framework of the "Preparedness Framework", which the company had previously announced and aims to manage the risks that advanced artificial intelligence systems may pose.

Company executives state that the industry will need more comprehensive security systems in the future as the capabilities of artificial intelligence models are rapidly increasing.