OpenAI partners with Poland’s NASK to strengthen AI safety

OpenAI and Poland’s NASK National Research Institute have signed an agreement to work together. Their main goals are to make artificial intelligence (AI) safer and more secure and to carry out technical research.

The partnership comes at an important time. Reports of AI agents acting outside their rules, escaping testing controls, and hacking government websites and companies are becoming more common at an alarming rate.

People have become more worried about OpenAI’s security measures after reports that a group of its AI agents broke out of their safety controls and carried out a cyberattack against Hugging Face, an AI platform.

NASK will serve as the Polish Artificial Intelligence Safety Institute. Through this partnership, the two organizations plan to test AI systems as they become more powerful, find weaknesses in their security, improve safety protections, identify content made by AI, and deal with problems involving AI that works in different languages.

The partnership also aims to build stronger connections between AI research organizations around the world. It will help OpenAI, the company behind ChatGPT, work with other research groups, including CAISSI in the United States and the UK AI Security Institute (AISI).

OpenAI says countries need shared international technical standards for advanced AI systems. These should include common ways to test what AI can do, check possible risks, test safety measures, and report problems or security incidents.

The partnership is also meant to help make sure that AI safety standards are followed around the world when AI is used in real-life situations.

In July, CERT Polska, a cybersecurity team that operates within NASK, used OpenAI’s AI models designed to help with cybersecurity work. This was part of an initiative called Daybreak.

Using OpenAI’s models, CERT Polska found six serious security weaknesses in MikroTik RouterOS. Some of these problems could have allowed attackers to take control of devices without logging in or proving their identity. After the weaknesses were discovered, the company that makes the software released fixes, helping prevent possible attacks.