OpenAI has classified its upcoming Astra model as its first “critical” cybersecurity model under the company’s Preparedness Framework, according to an announcement on X. The company claims Astra reached a significant threshold in cybersecurity capability during evaluation and is adding controls before continuing development.
OpenAI also states that it intends to make Astra broadly available and place its “advanced cyber capabilities” in the hands of defenders. The model has not been generally released.
In a follow-up post quoting OpenAI CEO Sam Altman, the company characterized Astra as “a powerful model” and maintained that restricting powerful models to a small group would not be a good strategy. The post added that Astra’s cyber capabilities require “a little bit longer” to deploy safely.
The term “critical” does not necessarily mean that Astra is considered dangerous in every use case. One Japanese-language account responding to the announcement clarified that the label refers to a serious category within OpenAI’s safety-evaluation framework, rather than a declaration that the model is inherently unsafe.
Source: X
