OpenAI has suspended development of a new AI model due to cybersecurity risks
In recent weeks, developers have reported several instances of AI-powered system breaches
OpenAI's new AI model, called Astra, has demonstrated dangerous capabilities in the field of cybersecurity / Photo: Stock all / Shutterstock.com
OpenAI, one of the leaders in the field of artificial intelligence, announced on August 7 that it had suspended some of its internal work on its next AI model, Astra. The model has demonstrated significantly greater capabilities in cybersecurity than expected. The company fears that the AI model could reach a “critical threshold” that would allow it to discover previously unknown vulnerabilities and develop solutions to exploit them without human intervention.
The developer of ChatGPT announced that it is strengthening security measures during the development and testing of new models. As part of these changes, OpenAI has suspended internal activities related to Astra that do not yet meet the new security control requirements. OpenAI also plans to collaborate with government agencies and organizations focused on AI safety to test Astra’s capabilities. A White House spokesperson stated that OpenAI has notified the administration of its plans to delay the release, according to Axios.
In addition, the company plans to develop guidelines for third-party partners on how to safely test more advanced models.
Context
OpenAI's decision came amid several incidents related to the autonomous capabilities of modern AI agents. Over the past two weeks, OpenAI and Anthropic have publicly acknowledged that their models unintentionally gained access to the computer systems of several organizations during testing, including Hugging Face, a platform for developing AI models. Meta also reported on August 6 that its recently released AI model, Muse Spark, had infiltrated a third-party organization’s computer system.
These cases demonstrate that AI agents are capable of independently performing actions that even vulnerability researchers cannot always predict in advance, according to Bloomberg. This underscores the need for more rigorous security testing of models and the use of secure environments for testing them, the agency concludes.
This article was AI-translated and verified by a human editor
