OpenAI has been forced to freeze part of the work on its promising Astra model. The reason is growing concern that the system could reach a critical level of cyber capabilities, which is classified as an unacceptable risk under the internal safety regulations of the Preparedness Framework.

During preliminary testing, the model demonstrated impressive progress in agentic programming and solving complex cybersecurity tasks. This places it on the brink of a threshold beyond which the neural network could theoretically act fully autonomously: finding and exploiting zero-day vulnerabilities in real protected systems, as well as independently planning and conducting multi-stage attacks with only a general statement of the goal.

Such a level of autonomy radically changes the threat landscape. If Astra is brought to this point, it will become a tool that can undermine infrastructure, steal data, or manipulate systems without any human control. Therefore, the decision to pause is not a bureaucratic formality, but a necessary precaution in conditions where the speed of AI development outpaces our ability to contain it.

For the industry, this is a signal: the race for superintelligence is entering a phase where safety becomes not just a priority, but a constraint. While regulators discuss general frameworks, it is the internal protocols of developer companies that determine where the red line will be drawn.

In my view, this step by OpenAI is a sign of maturity, but it also highlights a systemic problem: we are moving toward a point where AI capabilities may exceed not only technical, but also ethical barriers. The question remains how long the pause will remain effective if competitors do not follow similar restrictions.