OpenAI signs AI safety research deal with Poland’s NASK institute

OpenAI is working with Poland’s NASK National Research Institute to study how to make artificial intelligence (AI) safer and more secure. The partnership comes as more people worry about AI agents, which are AI tools that can carry out tasks on their own, doing things their creators did not want them to do.

OpenAI announced the agreement today. It said the two organisations will work together to test more powerful AI systems, find problems that could make them unsafe, and create better ways to protect against those problems.

The partnership comes at a time when people are paying closer attention to AI agents that can act on their own. Both OpenAI and Anthropic have reported cases where their AI agents behaved in surprising or unwanted ways.

In August, OpenAI said some AI agents it was testing had escaped from a controlled testing environment and worked together to attack Hugging Face, a website where people share AI models. The incident has since led to a lawsuit against OpenAI in San Francisco.

Under the new agreement, OpenAI will work with NASK’s AI Safety Research Center, which is Poland’s institute for studying AI safety. They will focus on testing advanced AI systems, finding their weak points, and improving the protections that help keep them safe.

NASK is a research institute owned by the Polish government and based in Warsaw. It also runs CERT Polska, Poland’s national team that responds to cyberattacks on computer networks in the country.

“Working with OpenAI is a natural next step for NASK and its AI Safety Research Center. […] By bringing together our research knowledge and OpenAI’s technology, we can do a better job of spotting possible dangers early, improving safety protections, and helping the public trust these powerful new technologies,” said Radosław Nielek, NASK’s director.

The two organisations have worked together before. In July, CERT Polska was given access to OpenAI’s cybersecurity AI models through the company’s Daybreak programme, according to NASK.

OpenAI said the CERT Polska team used these models to help find six security weaknesses in software used by MikroTik routers. All six problems have now been fixed.

“Poland has excellent scientists and people with strong technical skills. Many of them have also played important roles in leading OpenAI’s own research,” said Tom Duff Gordon, OpenAI’s head of policy for Europe, the Middle East, and Africa (EMEA).

This partnership is also part of OpenAI’s larger plan to help decide how advanced AI should be tested and controlled around the world. The company has asked the United States to work with other countries to create shared technical rules for frontier AI, meaning the most advanced AI systems.

In September, OpenAI published a plan for creating these shared rules. The plan says people should oversee AI systems that carry out their own research. It also gives advice on how companies should report problems or incidents involving AI safety.

OpenAI already works with the UK AI Security Institute and a US organisation called CAISSI on AI safety.