OpenAI's Astra model has been rated with Critical cyber capabilities, marking a significant advancement in AI security technology.
OpenAI has officially classified its upcoming AI model, Astra, as possessing a Critical level of cyber capabilities, a first for the company's AI systems, as of September 1. This designation is part of OpenAI's Preparedness Framework and pertains to a specific configuration known as Daybreak Blue, which will initially be available to a limited number of users following its launch.
Astra is designed to autonomously identify unknown vulnerabilities and create exploit methods for secure systems without needing detailed human guidance. This capability, previously only anticipated by OpenAI, has now been validated through testing.
Overview of Testing Results
To achieve a Critical designation within the Preparedness Framework, a model must autonomously discover and generate working exploits for zero-day vulnerabilities in various secure, real-world systems. The model must also be able to independently design and execute complex attacks based on general objectives.
In testing, Astra excelled, scoring a perfect 100% on the public ExploitBench. OpenAI developed an internal version of this benchmark due to concerns about using test data in training sets, which included 20 high-severity vulnerabilities in the V8 engine disclosed between June and August 2026. During evaluations, Astra successfully identified and exploited two unknown zero-day vulnerabilities as part of an exploit chain. OpenAI is currently notifying the developers of the affected software regarding these security flaws.
In further expert trials, Astra discovered unknown vulnerabilities within a secure browser and was able to create an attack chain, enabling it to escape a sandbox environment and execute commands on the main system after interacting with an HTML file. Additionally, Astra managed to escalate privileges from a standard user to root by combining various vulnerabilities within a secured operating system. It's important to note that the results reflect Astra's capabilities under the Daybreak Blue access level, not its standard configuration.
Access to Advanced Features Limited
OpenAI has recently delayed multiple phases of Astra's development and release while enhancing protections against potential cyber misuse and unauthorized actions. Following extensive testing, the company believes it is ready to launch the model soon, albeit with advanced features initially limited to a small group of testers before broader access becomes available through Daybreak Blue.
The model has been trained to reject requests for prohibited cyber assistance, adhering to strict limitations. In tests where attempts were made to bypass these restrictions, Astra declined 91.5% of such requests, a significant increase compared to the 59% refusal rate of GPT-5.6 Sol. OpenAI is also implementing additional monitoring systems to detect and automatically halt any unauthorized actions taken by the model.
Notably, in August, OpenAI slowed the development of advanced AI systems following an incident involving Hugging Face and Astra's preliminary evaluations.
