OpenAI and Poland’s NASK Partner on Frontier AI Safety: What It Means for Europe

How do you check whether an increasingly capable AI model is safe to use? That question is becoming harder for any one organisation to answer alone.

OpenAI and Poland’s NASK National Research Institute have signed an agreement to explore collaboration on AI safety, security and technical research. Announced by OpenAI on 23 October 2026, the partnership brings a frontier AI developer together with a Polish public research institute working on AI safety. [1]

The announcement describes areas for collaboration rather than completed research findings. It is an early step, but it offers a useful view of how AI developers and public institutions are beginning to work together on the risks of advanced systems.

What will OpenAI and NASK work on?

According to OpenAI, the partnership will involve NASK’s AI Safety Research Center, which serves as the Polish Artificial Intelligence Safety Institute.

The organisations plan to explore ways to evaluate increasingly capable AI systems, identify potential vulnerabilities and strengthen safeguards. OpenAI also says the agreement covers research into how advanced AI might contribute to a more secure digital environment. [1]

For readers, the distinction between a plan and a result matters. The announcement sets out the intended areas of work, but it does not establish that a particular model has passed an independent safety assessment.

Why frontier AI safety needs more than a benchmark

A model can perform well on a test and still behave unexpectedly when it interacts with software, websites or unfamiliar information.

That is why evaluating frontier AI involves more than comparing scores. Researchers may need to examine what tools a model can use, what information it can access, how it responds to misleading instructions and whether it can take actions without appropriate oversight.

OpenAI’s earlier guidance on safety cases for frontier AI training discusses alignment, containment and monitoring as parts of a broader safety argument. It also identifies operational measures such as reviewing safety claims and pausing training when significant problems are found. [2]

Those guidelines describe OpenAI’s approach to training; they are not proof that every risk has been eliminated or a description of a completed OpenAI–NASK evaluation.

A European perspective on advanced AI

The agreement is relevant to a wider conversation about frontier AI safety in Europe.

In October, the European Commission’s Scientific Panel of independent experts held a special meeting focused on frontier AI safety and risks. The Commission said the meeting would examine emerging risks and consider recommendations. [3]

The panel’s work and the OpenAI–NASK agreement are separate initiatives. Together, they illustrate why European institutions are paying attention to how advanced AI systems are evaluated and secured.

One question is how much independent access researchers need. A public report can explain what a company says it tested, but researchers may also need appropriate access to models, evidence and evaluation methods to assess claims for themselves.

The OpenAI announcement does not provide enough detail to determine the level of independent access NASK will receive or what findings will be made public.

What businesses should take from the announcement

Most organisations will not be training frontier models themselves. They may, however, use AI systems built on those models to search documents, write software, answer customers or operate internal tools.

Before giving an AI system access to sensitive information or the ability to take action, teams can ask:

What task is the system allowed to perform?

What information and tools can it access?

Who reviews important outputs or actions?

How will the organisation detect mistakes or misuse?

Can access be limited or suspended if something goes wrong?

These questions are a starting point for an organisation’s own risk assessment. They do not replace testing the specific product, version and deployment setting.

What to watch next

The most useful next update would be evidence of what the partnership produces: evaluation methods, research findings, identified limitations and an explanation of what can be independently reviewed.

For now, OpenAI and NASK have described an intention to collaborate. Whether that work leads to publicly available evidence remains to be seen.

As frontier AI becomes more capable, the important question is not simply whether a company has announced a safety partnership. It is what the partners learn, how they test their claims and how clearly they communicate the results.

Sources reviewed 27 October 2026. This article reports the organisations’ stated plans; Frontier AI has not independently evaluated the partnership or the models involved.