OpenAI приостановила разработку модели Astra из-за угроз кибербезопасности

OpenAI has frozen part of the development of its upcoming Astra model, citing cybersecurity concerns. The company did not specify which aspects have been paused or for how long.

What OpenAI Has Paused

OpenAI announced a temporary slowdown in work on a number of Astra components after an internal evaluation identified potential risks related to the model's cybersecurity capabilities. This is a partial, not a complete, freeze of the project.

• Astra development has been slowed, but not stopped entirely
• The reason is concerns about the model's cybersecurity capabilities
• OpenAI has not disclosed which specific features are subject to restrictions
• No public timeline for resuming full-scale development has been given

This is not the first time OpenAI has decided to slow development on safety grounds. Previously, the company delayed the release of certain model capabilities when internal reviews uncovered unwanted behavior prior to public release.

Why Cybersecurity Is a Special Risk Category

Cybersecurity is among the most sensitive areas of application for powerful language models. In theory, an advanced model could significantly lower the technical barrier for carrying out cyberattacks — which is precisely what security researchers are concerned about.

In its public documentation on Dangerous Capability Evaluations, OpenAI places cybersecurity in a critical risk category — on a par with biological and nuclear weapons. According to this framework, models that could meaningfully enhance the effectiveness of cyberattacks should not be made publicly available without additional safeguards.

The specific threat landscape that concerns researchers:

• Automated generation of functional malicious code
• Discovery of zero-day vulnerabilities in widely used software
• Creation of highly convincing phishing messages
• Planning attacks on critical infrastructure

How Pre-Release Safety Evaluation Works

Before every major release, labs such as OpenAI, Anthropic, and Google DeepMind conduct what is known as red-team testing: specialists deliberately attempt to use the model for destructive purposes. The goal is to identify risks before real users encounter them.

It was during such evaluations that Astra's concerning capabilities were discovered. As stated by OpenAI, the company chose to pause the relevant aspects of development rather than release them with unresolved risks.

Similar cases have occurred across the industry: various capabilities of flagship models have been delayed or restricted precisely after unwanted behavior was identified during internal testing — long before these decisions became public.

What This Means

OpenAI's public announcement of a slowdown in Astra's development on safety grounds is a signal of maturity in risk management. The company chose to openly communicate the restrictions rather than conceal them or release controversial features without proper scrutiny. For users awaiting Astra, this most likely means a delayed release — but also a higher level of confidence in the safety of the final product.