On September 3, OpenAI announced "GPT-6 Astra," a model capable of autonomously executing multiple steps within the field of cybersecurity. According to OpenAI, this is the first AI to receive a "critical" rating for its cybersecurity capabilities within OpenAI's internal evaluation framework. With the appropriate tools and access permissions, Astra possesses the ability to identify undiscovered vulnerabilities and develop exploits without human intervention. OpenAI acknowledged that it is difficult to track the model's inference process and that it may potentially evade monitoring; consequently, the company is strengthening defensive measures, including the isolation and monitoring of operational environments and the implementation of strict denial settings.
Source: OpenAIとGoogleは、ソフトウェアの脆弱性を独自に発見、検出、修正できるAIを発表した。 - Vietnam.vn (Google News: OpenAI)