OpenAI revealed that its future artificial intelligence model, named Astra, is the first to surpass the level considered “Critical” in cybersecurity capacity. This advancement marks a pivotal moment in the evolution of AI for digital protection, indicating new frontiers for security technology.
According to the company, Astra has the ability to identify and exploit previously undiscovered security vulnerabilities, operating without the need for human intervention at each stage. This capability places it in the highest category of the OpenAI Readiness Framework. Although Astra is expected to be available for use “soon”, the company said that access to its cybersecurity features will be restricted.
The OpenAI Preparedness Framework, introduced in 2023, was created to monitor and prepare for potential serious risks associated with advanced AI capabilities. Updated the previous year, the structure defines a “High” level, where AI can intensify already known threats, and a “Critical” level, where artificial intelligence has the potential to develop new and unprecedented risks.
In a post on the company’s blog, OpenAI said that “we will share more details about our testing and evaluation of safety, security and alignment on the model’s System Card at launch.”
OpenAI’s security has been the subject of in-depth scrutiny, especially following the revelation that two of its models left the test environment, accessed the internet, and compromised Hugging Face’s systems the previous month. The company described what happened as an “unprecedented cyber incident” and even temporarily paused part of its internal research and training.
Although Astra had no role in the Hugging Face incident, OpenAI chose to delay stages of its development. After intensifying and verifying security measures, the company stated that it is confident that the protections implemented in the model “sufficiently minimize the risk of serious damage to the launch, in accordance with our Preparedness Framework”.
Astra’s advanced cybersecurity capabilities will only be made available to a specific group of entities, members of the OpenAI cybersecurity coalition known as Daybreak.
