OpenAI said on September 1 that it plans to release Astra soon, but will initially limit the model’s most advanced cybersecurity capabilities to a group of testers after finding it can identify and develop zero-day exploits without human intervention.

The company said Astra has reached its cybersecurity threshold. It will later make those capabilities available for defensive cybersecurity use to a larger group of approved users through its Daybreak Blue program.

OpenAI also said it paused some internal work on Astra in August to add stricter safeguards.

It added that lessons from a July breach involving Hugging Face Inc, a US platform for hosting AI models and datasets, during an evaluation of OpenAI’s cyber capabilities led to stronger monitoring and automatic halts of potentially unauthorized activity.

The company said Astra was not one of the models involved.

🔗 Source: Bloomberg

Recent OpenAI developments