News
NASK and OpenAI Sign AI Safety Research Agreement
NASK, Poland's National Research Institute, and OpenAI have announced a partnership on safety research for advanced models. NASK representatives say it is the only such agreement in the European Union.
NASK and OpenAI have signed an agreement on AI safety research. It is meant to help find weaknesses in advanced models, test safeguards and identify new threats. The deal was announced on October 23, 2026.
What the agreement covers
The teams will study the capabilities and vulnerabilities of advanced models, as well as ways they can be misused. Other areas include detecting AI-generated content and assessing risks depending on language.
The agreement also supports broader work on shared AI safety standards: uniform methods for evaluating model capabilities, researching risks, testing safeguards and reporting incidents. The goal is to make results easier to compare and to respond to threats across borders.
AI safety cannot be ensured alone. - Radosław Nielek, director of the institute, NASK
Nielek added that, in his view, the agreement recognizes the competence of NASK's specialists. It will combine the institute's research experience with OpenAI's knowledge of how advanced models are developed, so that threats can be identified earlier and countered more effectively.
OpenAI's view
Poland has outstanding scientists and exceptional technology talent. - Tom Duff Gordon, vice president and head of public policy for EMEA, OpenAI
Duff Gordon stressed that as AI becomes more capable, an outside perspective is also needed to help verify assumptions, test systems and strengthen safeguards. OpenAI says joint tests with public institutes in the US and the UK helped uncover gaps in models and agents and improve protection. NASK is now joining these international efforts.
Earlier cooperation
In July, CERT Polska, part of NASK, gained access to OpenAI's advanced cybersecurity models through the Daybreak program. The team found six vulnerabilities in MikroTik RouterOS, some of which could be chained to take over a device without logging in. The manufacturer released patches, and CERT Polska confirmed they work.
Background and next steps
Evertiq, citing Reuters, notes that the agreement comes amid concern about autonomous AI agents. It mentions an incident in which a group of agents slipped out of control and attacked the Hugging Face platform. The sources give no timeline or budget for the joint research, so concrete results will take time.
