A new round of safety debates is brewing in the artificial intelligence industry: OpenAI has officially announced a partial suspension of work on its flagship model, Astra. The reason is not technical difficulties or a lack of computing resources, but the results of an internal risk assessment under the Preparedness Framework system. Based on my analysis, this is the first public case where a company voluntarily slows down development due to a potential threat in cyberspace.

Preliminary tests showed that Astra demonstrates significant progress in two key areas: agentic programming and solving cybersecurity tasks. This is not just an improvement in metrics—it is about approaching a critical threshold that OpenAI documents label as a "high level of cyber capabilities." This status implies that the neural network is capable, without human involvement, of finding and exploiting zero-day vulnerabilities in real protected systems, as well as independently planning and conducting complex attacks with only a general statement of the goal.

It is important to emphasize that this is not a hypothetical scenario, but actually measurable results. In my practice of analyzing AI models, such signals are usually ignored until an incident occurs, but here we see a preventive step. OpenAI, in essence, acknowledges that Astra could become a dual-use tool, where the potential for automating cyber defense is comparable to the risks to critical infrastructure.

It is not yet clear for how long the work has been frozen and which specific components of the model have been postponed. However, the very fact that the company applies its own safety regulations in real practice, rather than only on paper, is an important precedent for the entire industry. The question is whether other players are ready to follow this example, or whether the race for performance will prevail over caution. In my view, this is the right, albeit costly, step: it is better to lose a few months of development than to face the uncontrolled use of AI in cyberattacks.