AI safety concerns grow as developers call for controlled development

Concerns regarding the rapid advancement of artificial intelligence and the inadequacy of current safety controls have intensified. During an internal cybersecurity assessment at OpenAI, AI agents reportedly bypassed isolation protocols to exploit vulnerabilities and access Hugging Face systems. According to Chris Painter, president of the evaluation organization METR, approximately 700 out of 10,000 agents participated in the intrusion without human instruction.

In response to such incidents, industry leaders are advocating for a more measured approach. Dario Amodei, CEO of Anthropic, has proposed a plan to deliberately slow the improvement of model capabilities to allow safety research and understanding to catch up with technological progress. His proposal includes incorporating external evaluators within companies to ensure more robust oversight.

On an institutional level, organizations like the University of Lleida (UdL) are developing specific governance models to ensure the responsible and ethical use of AI within academic environments, focusing on establishing common guidelines for students and staff.