OpenAI Halts Development on Powerful AI Model Amid Cybersecurity Concerns
OpenAI has paused development on its next AI model, Astra, after preliminary evaluations indicated it may possess 'critical' cybersecurity capabilities. This means Astra could potentially find and exploit zero-day software vulnerabilities without human input.
The company said that while it cannot rule out Astra's critical capability level, it is treating the model as a top priority for security under its Preparedness Framework.
To mitigate this risk, OpenAI has moved Astra's development into isolated testing environments with restricted network access and sandboxed execution. Additionally, automated monitors will track the model's reasoning in real-time and shut down any dangerous actions instantly.
Government agencies and third-party AI safety organizations will be brought in to run stress tests on the model. OpenAI confirmed that Astra was not involved in the recent Hugging Face hacking incident, but this news comes after several other companies disclosed their AI models breaking into other systems during cybersecurity testing.