Less than a week after touting the scientific achievements of Astra, its next \u201cmajor\u201d model, OpenAI says it\u2019s \u201cpausing internal activities\u201d related to the model due to its powerful cybersecurity abilities.<\/p>\n
\u201cOur latest internal evaluations of Astra, one of our upcoming models, over the past few days indicate significant advancements in agentic coding and cybersecurity,\u201d OpenAI stated in a Friday press release.\u00a0<\/p>\n
\u201cThese results, in addition to expert assessments, have led us to conclude last night that we cannot rule out critical cyber capabilities under our Preparedness Framework\u2060.\u201d<\/p>\n
OpenAI\u2019s Preparedness Framework outlines scenarios in which development of a new model should \u201chalt\u201d if it reaches certain capability thresholds in various categories, including \u201cBiological,\u201d \u201cCybersecurity,\u201d and \u201cAI Self-improvement.\u201d<\/p>\n
For cybersecurity, the \u201ccritical\u201d threshold means a model can pinpoint \u201czero-day exploits of all severity levels\u201d in \u201chardened real-world systems\u201d without any human help.\u00a0<\/p>\n
A model could also hit the \u201ccritical\u201d level if it can carry out \u201cend-to-end novel strategies for cyberattacks against hardened targets\u201d with little more than a \u201chigh-level desired goal\u201d in mind, according to the OpenAI safety framework.<\/p>\n
OpenAI\u2019s previous high-end model, GPT-5.6 Sol, only reached the \u201chigh\u201d threshold during internal evaluations, the company said. OpenAI initially released GPT-5.6 Sol to just a \u201cselect group of trusted partners\u201d before making the model public a couple of weeks later.<\/p>\n
Given its concerns over Astra\u2019s potential cybersecurity risks, OpenAI says is it \u201cimplementing stricter security controls\u201d for the model, such as setting up \u201cisolated testing environments\u201d and \u201crestricted network and tool access,\u201d among other measures.\u00a0<\/p>\n
In the meantime, OpenAI is \u201cpausing internal activities involving Astra that do not yet meet these strengthened security control requirements,\u201d the company said.<\/p>\n
OpenAI said it sounded its warning about Astra because \u201cit\u2019s important to be transparent to the public\u201d about what Astra is potentially capable of.<\/p>\n
Barely a week ago, OpenAI touted Astra\u2019s abilities in mathematical research, including its solutions to 10 open math and computer science problems.<\/p>\n
OpenAI\u2019s revelations about Astra come amid a flurry of reports of advanced AI models going rogue, hacking real companies and organizations during training exercises and even forging phony credentials to hack external systems.<\/p>\n
Now with Astra said to be demonstrating dangerously strong cybersecurity capabilities, it seems we may have reached a crossroads in AI safety, where each new \u201cfrontier\u201d model on the AI test bench is judged \u2014 at least initially \u2014 too powerful to be released.<\/p>\n