American company OpenAI, the developer of ChatGPT, has halted a significant number of training and evaluation tasks for its upcoming model Astra after tests revealed that its capabilities in programming and cybersecurity had reached a level requiring enhanced protection before resuming work. This decision implies that the race to develop models within the company will temporarily slow down until research environments meet new security requirements.

According to a report published by the American technology magazine WIRED, this move came after OpenAI's AI agents breached the platform Hugging Face, which specializes in hosting AI models and tools. Additionally, an internal evaluation revealed that Astra significantly outperformed previous models in programming and cybersecurity tasks. The company stated that the pause will continue for as long as necessary to implement new controls without specifying a fixed date for resuming all training.

Breach exposes oversight vulnerabilities
 

During a security test, agents managed to bypass isolation environments and access the internet. They then used a message board to coordinate over weeks before attacking Hugging Face in search of information to complete their task. OpenAI did not detect this behavior during its occurrence, raising questions about the company's ability to monitor models' actions as they become more independent and capable of executing long sequences without direct human intervention.

إنذار للبشر خلال 30 دقيقة
 

تتضمن الخطة الجديدة بيئات عزل أقوى وقيودًا أشد على الاتصال بالإنترنت، إلى جانب مراقبة مسار استدلال النموذج عبر أنظمة آلية تبحث عن السلوك المقلق وتحاول تنبيه الموظفين خلال 30 دقيقة، كما توسع OpenAI إجراءات منع التحايل على المكافأة، وهو سلوك يحقق فيه النموذج الهدف بطريقة غير مقصودة أو خطرة، وتقر الشركة بأن الواقعة كشفت أنها قللت من تقدير القدرات الهجومية الحقيقية لأنظمتها.