Explainer — OpenAI said on October 23 that it had signed an agreement with Poland’s NASK National Research Institute to explore cooperation on AI safety, security and technical research. NASK’s AI Safety Research Center serves as Poland’s Artificial Intelligence Safety Institute. The announcement names three broad goals: evaluating increasingly capable systems, identifying vulnerabilities and strengthening safeguards. It does not disclose a budget, timetable, named benchmark suite or a promise that results will be published, so this is best understood as a framework for future work rather than a finished testing program.
The practical issue behind the agreement is access. Independent researchers can study public model behavior, but the most useful safety work may require controlled access to pre-release systems, internal tooling or detailed technical evidence. OpenAI says its work with public research bodies in the United States, the United Kingdom and elsewhere has already helped identify and fix security weaknesses in AI agents. A national institute can also contribute local-language testing and knowledge of the country’s public-service and information-security environment—areas that a single global test set may miss.
The timing matters in Europe. The European Commission says its AI Office develops tools, methodologies and benchmarks for evaluating general-purpose AI models and for identifying models that may pose systemic risk. Commission guidance also states that providers of systemic-risk models must evaluate and mitigate risks, report serious incidents and maintain adequate cybersecurity. The Commission’s enforcement powers for these general-purpose AI obligations began applying on August 2, 2026.
That makes the OpenAI–NASK relationship potentially relevant beyond voluntary safety research. A technically capable national partner could generate evidence that informs European evaluation practice or exposes weaknesses before deployment. That is an inference, not a commitment announced by either party. The agreement’s language is deliberately exploratory, and cooperation with a model developer should not be confused with an independent audit, regulatory approval or proof that a system is safe.
The useful way to judge the pact is therefore to watch for deliverables: jointly defined evaluations, transparent methods, published findings where security permits, clear rules for researcher independence, and evidence that discovered vulnerabilities lead to fixes. If those artifacts appear, the partnership could help turn abstract international coordination into repeatable technical work. If they do not, the announcement will remain a statement of intent rather than a measurable addition to Europe’s AI-safety capacity.