KOSUNI CRYPTO
Korea's Crypto Pulse, in English
$BTC $ETH $XRP $SOL $DOGE
News
Neutral

OpenAI Confirms Astra Model Reaches Critical Security Threshold

Published September 26, 2026 9:03 PM · 7 views
OpenAI Confirms Astra Model Reaches Critical Security Threshold

OpenAI confirmed that its next-generation model Astra has reached the 'Critical' cybersecurity capability threshold within its prepared framework. Astra is the first model designated with this rating by OpenAI. In a post dated August 1, OpenAI stated that Astra can identify and develop methods to exploit unknown security flaws without human step-by-step instructions, provided it has appropriate tools and access permissions.

Astra achieved a perfect score of 100% on ExploitBench, which evaluates models' ability to develop attacks based on known vulnerabilities. During internal testing using 20 high-risk V8 vulnerabilities, Astra demonstrated higher arbitrary code execution rates with fewer output tokens compared to GPT-5.6 Sol. The model also discovered and utilized two previously unknown zero-day vulnerabilities during the evaluation process, which OpenAI is now disclosing to the responsible parties.

OpenAI clarified that these results were obtained under the 'Daybreak Blue' access condition, featuring advanced cybersecurity capabilities rather than the standard user version of Astra. Additionally, OpenAI resumed large-scale reinforcement learning (RL) training on August 28 following a pause after the Hugging Face security incident, though some experimental training remains on hold.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 오픈AI, 아스트라 보안 최고위 임계점 도달 확인