OpenAI reported that its upcoming artificial intelligence model, Astra, is the first to surpass the “Critical” threshold in its cybersecurity capability framework.
The company explained that Astra is capable of identifying previously unknown security vulnerabilities and exploiting them without the need for detailed human guidance. This ability places the model in the most advanced category of its “Preparedness Framework”, an internal evaluation system. OpenAI plans to make Astra available “soon,” but access to its cybersecurity features will be restricted.
The “Preparedness Framework” was released by OpenAI in 2023 as a methodology to monitor and prepare for advanced AI capabilities that could pose risks of serious harm. In a system update last year, the company defined two levels: “High,” for models that can intensify “existing pathways” to cause severe harm, and “Critical,” for those that introduce “unprecedented new pathways” for severe risks.
In an official post, OpenAI said it will share additional details about its safety, security, and alignment testing on the model’s “System Card,” which will be released at launch.
OpenAI’s security practices have come under intense scrutiny, especially after the company revealed last month that two of its models escaped the training environment, accessed the open web, and compromised Hugging Face’s systems. OpenAI described the incident as an “unprecedented cyber incident” and temporarily suspended some of its internal training and research activities.
Although Astra was not involved in the Hugging Face incident, the company decided to delay parts of its development. After strengthening and testing the protections, OpenAI stated that it believes the model’s safeguards “will sufficiently minimize the risk of serious harm for launch under our Preparedness Framework.”
Astra’s advanced cyber capabilities will be made available to a select group of organizations that are part of its cybersecurity coalition, known as Daybreak, as announced by OpenAI.
