OpenAI Signs AI Safety Research Agreement With Poland's NASK

OpenAI said on October 23 that it has signed an agreement with NASK, Poland's National Research Institute, to explore joint work on AI safety, security and technical research. The deal connects OpenAI to NASK's AI Safety Research Center, which the company says serves as Poland's Artificial Intelligence Safety Institute. What the agreement obligates either side to actually do — in budget, staffing, model access or publication terms — was not disclosed.

What the agreement says

According to OpenAI, the two organizations will collaborate on evaluating increasingly capable AI systems, identifying potential vulnerabilities and strengthening safeguards. The framing in the announcement is deliberately open-ended: the parties have agreed to "explore collaboration," which is the language of a memorandum of understanding rather than a contract with deliverables.

OpenAI positions the NASK deal alongside its existing relationships with the US Center for Advancing Standards and Innovation for Super Intelligence (CAISSI) and the UK AI Security Institute. The company says that body of work has already helped strengthen safeguards, find and fix security flaws in AI agents, and advance frontier safety research. It does not break out which results came from which partner, so there is no public basis for comparing how much any individual institute has contributed.

Radosław Nielek, director of NASK, said in a statement provided by OpenAI that the institute's safety center works on safeguards against emerging risks, detection of synthetic AI-generated content, malicious uses of that content, and the multilingual dimension of those problems. He argued that safety cannot be achieved by frontier developers acting alone and that cooperation between public safety institutes and labs is necessary for AI to be reliable in practice.

Tom Duff Gordon, OpenAI's VP and head of EMEA policy, said Poland has produced scientists and engineers who have led work inside OpenAI, and that outside expertise is needed to challenge the company's assumptions and test its systems. That is a reasonable statement of intent, though it leaves the key operational question open: an external institute can only challenge assumptions if it has meaningful access to models, internal evaluations and the reasoning behind deployment decisions. The announcement does not describe what access NASK will receive.

The one concrete precedent: CERT Polska and MikroTik

The most specific item in the announcement is not the new agreement but what came before it. OpenAI says that in July, CERT Polska — the national computer emergency response team, which operates inside NASK — gained access to the company's cyber-capable models through a program called Daybreak. Using those models, the team found six vulnerabilities in MikroTik RouterOS, including flaws that could be chained together to take over affected devices without authentication. OpenAI says the vendor shipped fixes and that CERT Polska confirmed the fixes stopped the attacks it had been observing.

That is a useful data point, and unusually falsifiable for this genre of announcement: a named product, a count of findings, a patch, and a confirmation from the defender. It is also narrower than it may first appear. The announcement does not say how much of the discovery work the models did versus the human analysts, how many false leads the team worked through, or how the six findings compare with what CERT Polska would have found using conventional tooling over the same period. Without that baseline, the result shows that the models were useful in a real workflow, not how much they changed the economics of vulnerability research.

The broader tension is familiar to anyone following offensive security. Models that are good enough to find chainable authentication bypasses in widely deployed network gear are, by construction, useful to attackers. OpenAI's approach here is gated distribution to vetted defenders — the Daybreak framing — which is a defensible policy choice but one whose durability depends on how tightly access can be controlled as similar capabilities appear in open-weight models.

Why national institutes, and what they are supposed to add

OpenAI used the announcement to repeat its call for the United States to work with other countries on international technical standards for frontier AI, drawing on safety institutes and standards bodies. The argument is that shared methods for evaluating capabilities, assessing risk, testing safeguards and reporting incidents would let countries compare evidence and respond to cross-border risks. The company adds that without national partners like NASK, global standards risk drifting away from how AI is actually used on the ground.

There is substance to that argument beyond the diplomacy. NASK's stated specialties — synthetic content detection and multilingual risk — point at a real gap in frontier safety evaluation, which has been heavily weighted toward English. Safeguards that hold in English can degrade in lower-resource languages, and a Polish institute testing Polish-language failure modes is the kind of work a US or UK lab is poorly placed to do on its own. NASK also runs operational infrastructure through CERT Polska, which gives it exposure to live incidents rather than lab benchmarks.

The structural weakness is independence. A safety institute that depends on a lab for model access, compute and technical support is not an auditor in any strong sense. The announcement does not say whether NASK will be free to publish findings OpenAI dislikes, whether evaluations will happen before or after deployment, or whether results will be made public at all. Those are the details that determine whether this kind of arrangement functions as external scrutiny or as collaborative engineering with a government co-signer.

What is not in the announcement

No funding figures, no duration, no headcount, no list of models in scope, and no first project. There is no indication of how this relates to European regulatory obligations, and no statement about whether NASK's findings feed into any formal process. The agreement is described as exploratory, which means the realistic near-term output is a scoping exercise rather than published research.

For builders, the practical read is modest: another national institute is now inside the tent, and the one measurable result so far came from defensive security work that predates the agreement. Whether the partnership produces evaluation methods others can reuse — particularly for non-English safeguards — is the test worth applying in a year.

What to watch

• Whether NASK publishes evaluation results independently, and whether it retains the right to publish findings unfavorable to OpenAI.
• Whether the collaboration covers pre-deployment testing of frontier models or only post-release analysis; the announcement does not say.
• Whether multilingual and synthetic-content evaluations produced in Poland get adopted as shared methods by other safety institutes, or stay national.
• How Daybreak access is governed as more defenders request cyber-capable models, and whether OpenAI discloses the criteria for admission.
• Whether follow-on results are reported with baselines — time spent, false positives, comparison to existing tooling — rather than raw counts of vulnerabilities found.