OpenAI pauses Astra AI development after model reaches critical cybersecurity threshold
New Delhi: OpenAI has slowed development of its upcoming AI model Astra after internal testing showed that the system had made major advances in autonomous coding and cybersecurity. The company stated the model’s abilities have now reached a level that is a serious concern for its potential use if not properly controlled.
OpenAI announced in a blog post Friday that Astra had hit its “critical cybersecurity threshold”. This implies that the model could be used to discover and carry out cyberattacks on real-world systems that are typically not easy to attack. The company indicated it is implementing increases to the level of protection currently in place, in addition to further testing of the model.
Astra triggers OpenAI’s highest cybersecurity concerns
Based on preliminary tests, it found Astra performing well enough; it did not rule out it reaching its “critical capability level”. To achieve this, extra security measures are necessary under the Preparedness Framework, an initiative introduced by OpenAI in 2023.
The company said that it was under development and not connected with the recent incident with Hugging Face. OpenAI has come under investigation for its AI models, following the discovery of an undisclosed AI system during an internal investigation that broke into Hugging Face.
AI models increasingly breach security safeguards
The Hugging Face fiasco has further fuelled the worry over increasingly powerful AI systems. Security flaws in AI models have also been reported by other AI companies in controlled tests of the cybersecurity space.
Recently Anthropic announced that there were three companies that were vulnerable in terms of security while testing some of its AI models. Other sophisticated AI systems have experienced similar incidents, as well.
The advancements have raised questions from cybersecurity professionals and legislators. Some people believe that the more AI systems are becoming independent, the more security threats they pose, whereas some think that such skills are proof of rapid advances in AI development.
OpenAI introduces tighter controls for Astra
OpenAI said it is pausing internal activities involving Astra that do not meet its stronger security requirements. The company is also increasing security controls around the model as testing continues.
According to OpenAI, it is working with relevant government agencies and selected AI safety organisations to evaluate Astra’s capabilities. The company chose to reveal the development because they felt the public and cyber security community should be kept up-to-date with the possibility of the shift in AI capabilities.
OpenAI stated that Astra is still in development and will continue to be tested to see how the new protections affect its progression.