OpenAI pumps the brakes on new Astra mannequin over cybersecurity issues
“These outcomes, along with professional assessments, have led us to conclude final night time that we can not rule out important cyber capabilities below our Preparedness Framework.”
OpenAI’s Preparedness Framework outlines situations during which improvement of a brand new mannequin ought to “halt” if it reaches sure functionality thresholds in numerous classes, together with “Organic,” “Cybersecurity,” and “AI Self-improvement.”
For cybersecurity, the “important” threshold means a mannequin can pinpoint “zero-day exploits of all severity ranges” in “hardened real-world techniques” with none human assist.
A mannequin may additionally hit the “important” degree if it may perform “end-to-end novel methods for cyberattacks in opposition to hardened targets” with little greater than a “high-level desired objective” in thoughts, in keeping with the OpenAI security framework.
OpenAI’s earlier high-end mannequin, GPT-5.6 Sol, solely reached the “excessive” threshold throughout inner evaluations, the corporate stated. OpenAI initially launched GPT-5.6 Sol to only a “choose group of trusted companions” earlier than making the mannequin public a few weeks later.
Given its issues over Astra’s potential cybersecurity dangers, OpenAI says is it “implementing stricter safety controls” for the mannequin, akin to organising “remoted testing environments” and “restricted community and power entry,” amongst different measures.
