OpenAI’s new Astra model may pose critical cybersecurity risks

OpenAI disclosed that one of its unreleased large language models, known as Astra, may pose a significant cybersecurity risk. The algorithm, which recently solved 10 long-running math problems, could qualify for a critical designation under OpenAI's Preparedness Framework, as reported by Silicon Angle.

Astra is the first of OpenAI's large language models to potentially receive a critical cybersecurity risk rating. This designation is based on recent tests indicating the model might be capable of finding zero-day exploits in hardened systems without human assistance or launching cyberattacks based on high-level goals. OpenAI is mitigating these risks by restricting Astra's access to the public web and developing it within sandboxed environments with limited network and tool permissions. The company is also enhancing security measures to prevent the theft of Astra's code, focusing on the encryption of its weights.

Internal AI agents powered by Astra are monitored for malicious activity by analyzing their chain of thought. OpenAI plans to share its cybersecurity workflows with third-party testing partners, government agencies, and AI safety organizations.

Source: Silicon Angle