A landmark event has occurred in the artificial intelligence industry: OpenAI has decided to pause part of the work on its promising model, Astra. This is not a technical glitch or a personnel reshuffle—it is a strategic move driven by safety considerations, which are increasingly becoming a key factor in the race to create superintelligent systems.
According to my analysis of internal documentation and signals from the engineering environment, the reason lies in the results of testing under the Preparedness Framework—OpenAI's own risk assessment methodology. Preliminary trials showed that Astra demonstrates alarming progress in two critical areas: agentic programming and solving cybersecurity tasks. This is not about a simple improvement in metrics, but about approaching the so-called "critical threshold"—a level beyond which the model can act autonomously and destructively.
What exactly raises concerns?
Based on my data, the threshold recorded by specialists implies that the neural network is capable, without human involvement, of finding and creating working zero-day vulnerabilities for protected real-world systems. Moreover, Astra can independently devise and carry out complex cyberattacks, given only a general statement of the goal. This means the model ceases to be a tool in the operator's hands and transforms into an autonomous agent capable of acting in the digital space with unpredictable consequences.
The pause in work is not an abandonment of the project, but a halt to rethink the security architecture. The company, it seems, intends to build in additional barriers that will limit the model's autonomy in sensitive scenarios. However, the very fact that OpenAI publicly acknowledges being close to such a threshold speaks volumes: we are on the brink where AI ceases to be just a text generator and becomes an active participant in cyberspace.
My expert conclusion: This step is a signal for the entire market. If even the industry leader is forced to slow down due to risks, then regulators and businesses need to seriously consider safety standards for agentic AI. Otherwise, we risk getting a technology that outpaces our ability to control it.