OpenAI's Next AI Model Astra...in its Preparedness Framework Critical It was officially rated at that level. This is the first time OpenAI has assigned this rating to an actual model.
The key point of this announcement isn't simply that coding performance has improved. Given the right tools and access privileges, Astra can find unknown vulnerabilities in well-protected systems, and It was assessed that it had reached a level where it could plan an attack without step-by-step instructions from a human.
However, Astra hasn't been fully released to the general public yet. OpenAI has strengthened its safety measures ahead of the launch and announced that access to advanced cyber capabilities will be limited from the outset.
| From “Possible” to an official “Critical” rating
As recently as early August 2026, OpenAI released a preliminary assessment stating that it could not rule out the possibility that Astra had reached the “Critical” level. After gathering additional evidence and conducting automated benchmarks and expert evaluations, the company concluded that, as of September 1, Astra met the threshold.
The reason this change is important is that safety standards are also being raised. It’s not enough to simply classify them as high-performance models; stronger security controls and safeguards against misuse are required throughout the development process and before and after release.
| What "Critical Cyber Capabilities" Means
In the Preparedness Framework, a "Critical" rating is different from simply answering general security questions or describing known vulnerabilities. The criteria are whether you can identify new vulnerabilities in various hardened real-world systems and create viable attack vectors, or whether you can execute a new attack strategy from start to finish even when given only high-level targets.
OpenAI explained that Astra demonstrated this capability under conditions where it had the right tools and access privileges. Therefore, this finding does not mean that the model produces the same results in every environment; rather, it indicates that the model was found to possess capabilities that could lead to serious harm under specific evaluation conditions.
| Performance Gaps Identified in the Evaluation
According to OpenAI, Astra excels in its ability to discover vulnerabilities and develop attack code, GPT-5.6 SolIt outperformed the competition by a wide margin and required fewer tokens to perform the same task. ExploitBench, which evaluates the likelihood of an attack based on known vulnerabilities, reported that it scored 100%.
Even in an internal assessment consisting of 20 recently disclosed high-risk vulnerabilities, Astra showed a higher success rate for arbitrary code execution than GPT-5.6 Sol. During the evaluation process, two previously unknown vulnerabilities were discovered, and OpenAI is in the process of notifying the software’s administrators of these issues.
These figures are based on results obtained under the evaluation environment and advanced defense conditions established by OpenAI. You shouldn't assume that this is equivalent to the performance of the default product settings that general users will receive.
| Reasons for Slowing Down Development and Strengthening Safety Measures
OpenAI announced that, over the past few weeks, it has delayed some of Astra’s development and release schedule while testing safeguards to prevent cyber misuse and unauthorized model behavior. It has also put some large-scale reinforcement learning projects on hold until they meet security requirements.
The new safety measures include training to more reliably reject dangerous requests, a system-level safety classifier, abuse detection, and monitoring to halt unauthorized actions. It's a structure that doesn't rely solely on the model's own resistance capabilities but instead layers multiple layers of defense.
| The release is scheduled, but advanced features will be limited
OpenAI announced that it plans to release Astra soon, but did not provide a specific launch date. Additionally, the highest level of cybersecurity features will not be available to all users from the start.
Initially, a limited group of evaluators will use the advanced features, and later, the program—which is designed for defensive purposes— Daybreak BlueIt has been announced that the scope of access will be expanded through this method. Upon the official launch, we plan to disclose more detailed information on safety, security, and compliance assessments along with the system cards.
Key Points to Check Going Forward
This announcement signals that AI cyber capabilities have moved beyond the hypothetical stage of research to the point where they are shaping actual deployment policies. In the future, how other AI companies evaluate and control models with similar capabilities could become the industry-wide standard.
However, most of the results currently available are based on OpenAI's own evaluations. To more accurately assess the effectiveness of safety measures, information must be disclosed regarding the official system specifications, independent third-party evaluations, actual conditions for accessing advanced features, and the levels of false alarms and disruption to normal operations following the system’s launch.
References
Tags
#OpenAI #Astra #Artificial Intelligence #AI Model #Cybersecurity #PreparednessFramework #AI Safety #Vulnerability Analysis #Security Model #DaybreakBlue #AI Governance #AI News
