Odaily Planet Daily reports that OpenAI has stated it cannot rule out that its upcoming model, Astra, possesses "critical" cyber capabilities, prompting OpenAI to expand its security testing and suspend internal activities that do not meet stricter safety requirements. This is the latest indication of the rapid advancement in the cyber capabilities of AI models.
OpenAI stated that, after internal evaluations of one of its upcoming models, Astra, “we cannot rule out the possibility that it possesses critical cyber capabilities.” Before releasing Astra, OpenAI will expand testing and safety measures for the model and slow its development in accordance with its “Preparedness Framework,” first published in 2023, until appropriate safeguards are in place.
OpenAI also stated that Astra was not involved in the Hugging Face vulnerability attack. This may be the first instance of a leading AI laboratory committing to slow down its AI model development due to cybersecurity concerns. (Axios)