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

OpenAI's Astra AI Model Finds Unknown Vulnerabilities Without Human Help

Published September 26, 2026 7:39 PM · 8 views
OpenAI's Astra AI Model Finds Unknown Vulnerabilities Without Human Help

OpenAI's new AI model Astra has reached a level where it can identify unknown security vulnerabilities and link them to attack methods without step-by-step human assistance. OpenAI announced on January 1 that Astra is the first model to meet its 'Critical' cybersecurity capability standard under its Preparedness Framework.

The framework assesses whether a model can create attack tools using zero-day vulnerabilities in enhanced real-world systems or develop and execute high-level cyberattack strategies. OpenAI stated that Astra scored 100% on ExploitBench, an evaluation measuring the ability to craft attack code from known vulnerabilities. In an internal assessment using 20 high-risk V8 vulnerabilities released between June and August 2026, Astra discovered and utilized two unknown vulnerabilities while building attack chains.

CoinDesk reported that Astra bypassed enhanced browser sandboxes to execute commands on the host computer and linked multiple operating system vulnerabilities to gain root access. OpenAI delayed parts of Astra's development and release to strengthen protective measures and plans to initially offer advanced cybersecurity features only to select alpha testers.

Korean Source

This article is an English localization of a Korean-language crypto news report. Original headline: 오픈AI 아스트라, 인간 도움 없이 공격 구성