OpenAI has partnered with Poland’s NASK National Research Institute to advance artificial intelligence safety. The agreement, announced on October 23, 2026, focuses on evaluating advanced AI systems and identifying security weaknesses. The two organisations also aim to develop stronger safeguards as AI capabilities continue to grow.
The partnership brings together OpenAI’s technical expertise and NASK’s research capabilities. Their work will address emerging risks and help researchers assess the reliability of increasingly capable AI systems.
What the OpenAI–NASK partnership will investigate
OpenAI will collaborate with NASK’s AI Safety Research Center, which serves as Poland’s Artificial Intelligence Safety Institute. The teams will explore methods for evaluating advanced AI systems and detecting security vulnerabilities.
Researchers will also examine ways to strengthen protective measures. Their work will consider emerging AI risks, synthetic content detection and the potential misuse of artificial intelligence.
OpenAI believes collaboration with public research institutions can improve AI safety. Independent researchers can challenge existing assumptions and test systems from different perspectives. Their findings may help developers improve safeguards as AI capabilities advance.
Why AI agent safety matters
AI agents can now perform increasingly complex tasks with less direct human supervision. They can assist with research, software development and automated workflows. However, greater autonomy also creates new security challenges.
Researchers must ensure these systems follow their intended instructions and respect safety limits. They also need reliable methods to detect unexpected behaviour before it causes harm.
Recent evaluations have highlighted weaknesses in AI monitoring systems. These findings underline the importance of testing safeguards under realistic conditions. Developers must also ensure that monitoring tools can identify potentially dangerous actions.
The OpenAI–NASK collaboration adds to broader efforts to improve AI security. Independent evaluations could help researchers identify weaknesses before attackers exploit them.
Poland’s existing cybersecurity work with OpenAI
The new agreement builds on earlier cooperation between OpenAI and Polish cybersecurity experts. OpenAI previously provided cyber-capable models to CERT Polska through its Daybreak programme in July.
According to OpenAI, the team used these models to identify six vulnerabilities in MikroTik RouterOS. The company reported that some flaws could be combined to take control of affected devices without authentication.
The vendor subsequently released security fixes. CERT Polska confirmed that the updates prevented the attacks its team had observed.
This work demonstrates how AI models can support defensive cybersecurity research. Specialists can use them to investigate software weaknesses and identify potential attack paths. However, human expertise remains essential when validating findings and prioritising fixes.
Could the partnership influence AI safety standards?
OpenAI has advocated international cooperation on technical standards for advanced AI systems. Common evaluation methods could help researchers compare results and assess potential risks more consistently.
Such approaches could also improve safeguards and encourage better reporting of significant incidents. Cooperation between AI developers and independent institutions may help establish more reliable testing practices.
The partnership could contribute Polish research expertise to these wider efforts. However, the announcement does not establish a new international regulatory framework. Nor does it guarantee that researchers will eliminate the risks associated with advanced AI.
The practical impact will depend on the research the teams conduct and the findings they produce. Developers must then use those findings to improve their systems and protective measures.
As AI capabilities grow, safety research will become increasingly important. Partnerships between technology companies, independent researchers and public institutions can help identify weaknesses and develop more reliable safeguards.
Read more : AI agents disabling their own safety monitors
