OpenAI Slows Astra after Model Breaks Cyber Security Threshold

Illustration of OpenAI. The company temporarily halted some work on the Astra model after its cyber capabilities triggered additional security. Illustration photo: Unplash.

JAKARTA - OpenAI has temporarily halted some work on the Astra AI model after an internal evaluation found that its cyber capabilities were evolving to trigger additional security.

TechCrunch, quoted Saturday, August 8, reported that Astra, which is still in the development stage, showed great progress in agentic coding, the ability of AI to run a series of programming tasks more independently, as well as cybersecurity.

OpenAI said the model had reached a cyber security tipping point.

According to the company, the threshold means that the model can independently identify and execute cyber attacks on real systems that are usually well protected.

The findings triggered the implementation of additional security based on the OpenAI Preparedness Framework, which was created in 2023.

"As we continue to test and evaluate this model, our initial evaluation shows a strong enough performance that for now we cannot rule out a level of Critical ability," OpenAI wrote.

This means that OpenAI has not confirmed that Astra has reached the Critical level of ability. Initial evaluations only show that the possibility cannot be ruled out.

The company also emphasized that Astra was not involved in the exploitation of Hugging Face.

The announcement comes after OpenAI was previously highlighted due to another unreleased model breaking into the Hugging Face system in internal testing.

TechCrunch called the incident the first verifiable case of an AI lab losing control of its model.

Since then, OpenAI and other AI labs such as Anthropic have also revealed a number of incidents when AI models break into isolated testing environments or sandboxes and pose a threat during cyber security tests.

A sandbox is a testing environment that is separated from the main system to limit risks when models or software are tested.

The series of events triggered a variety of responses from cybersecurity experts, policymakers, and AI companies. Some are asking for tighter oversight.

On the other hand, there are also those who see such capabilities as a technological achievement that shows the progress of the AI model.

OpenAI said the information about Astra was shared because the company believed the public and the safety and security community needed to know about the changes to the model's capabilities.

The company is now tightening security controls and temporarily suspending internal activities involving Astra if it does not meet new security standards.

OpenAI is also working with relevant government agencies and a number of selected AI safety organizations to further test Astra's capabilities.