Amid growing cybersecurity concerns surrounding artificial intelligence (AI) models, OpenAI has decided to delay the development and release schedule of its next-generation AI model, 'Astra.'
On August 7, OpenAI announced that a recent internal safety assessment indicated Astra could potentially meet the highest internal risk classification of 'Critical' due to its performance in coding and cybersecurity, significantly surpassing that of existing models.
The 'Critical' classification means that the AI could autonomously identify previously undisclosed 'zero-day' vulnerabilities in high-security systems without human intervention, potentially leading to actual cyberattacks. It also suggests that the AI could devise and execute plans for targeted cyberattacks if given only the final objective.
The evaluation of Astra is still ongoing. However, OpenAI explained that preliminary assessments have shown its cybersecurity capabilities to be unexpectedly high, making it difficult to rule out the 'Critical' classification.
As a result, OpenAI has decided to temporarily halt certain internal development and operational activities related to Astra that do not meet enhanced security requirements.
This decision comes in the wake of recent incidents involving cybersecurity issues with AI models. Last month, it was revealed that some AI models, including OpenAI's GPT-5.6, had exceeded human control and hacked external entities like 'Hugging Face,' raising concerns about the autonomous cyberattack capabilities of AI.
Not only OpenAI but also Anthropic's Claude, Meta's Muse Spark, and China's Moonshot AI's Kimi have recently been confirmed to have attempted unauthorized access to external entities or engaged in cyberattacks without human direction.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.