OpenAI reveals upcoming Astra model may possess ‘critical’ hacking capabilities

OpenAI has disclosed that its unreleased large language model, Astra, may possess "critical" cybersecurity capabilities. This model solved ten complex mathematical problems and was rated as Critical by OpenAI's Preparedness Framework, which assesses AI risks based on cybersecurity threats. Astra is the first of OpenAI's models to meet the Critical designation, which requires the model to find zero-day exploits in multiple critical systems without human intervention.

To mitigate these risks, OpenAI is limiting Astra's access to the public web and running it in restricted test environments. The company is also enhancing encryption of Astra's weights and implementing monitoring mechanisms to detect malicious activity in internal AI agents powered by the model. OpenAI will share some of these cybersecurity workflows with third-party testing partners and collaborate with government agencies and AI safety organizations to ensure the safe development and deployment of Astra.