Futureweb

Artificial Intelligence

OpenAI and Poland's NASK signed a research agreement. The agreement is not a new model release.

OpenAI and Poland's NASK signed an October 23 research agreement. It is not a model release. Reuters adds context the posts do not make a term of the deal.

The short version

OpenAI says it signed an agreement with NASK to collaborate on AI safety research. NASK's center would help evaluate systems and safeguards. The posts release no model. Reuters reports the signing and adds other safety disputes. Those disputes are Reuters's frame, not a term of the agreement.

An agreement to look together

OpenAI's October 23, 2026 post says it signed an agreement with NASK to explore collaboration on AI safety, security, and technical research. NASK's AI Safety Research Center, which the post says serves as Poland's AI safety institute, would work with OpenAI on evaluating more capable systems, identifying vulnerabilities, and strengthening safeguards. NASK published its own announcement of the partnership. Reuters carried the signing the same morning. The verb in the primary post is explore and collaborate. There is no model card.Model Garden is a different OpenAI document.

What the quotes commit them to

Nielek, in OpenAI's rendering, talks about safeguards against emerging risks, detecting synthetic content, and the multilingual side of that problem, and he says safety will not be achieved by a developer alone. Duff Gordon says Poland has researchers who already contributed to OpenAI's work, and that the agreement lets the company test its assumptions with people outside the building. Those are ambitions for a joint research program. An ambition in a quote is not a delivered evaluation. This page does not have a joint paper, a shared test set, or a list of systems NASK has already scored under the new agreement.Craft is a product launch, not this.

July stays in July

OpenAI says CERT Polska, inside NASK, received access in July through a program it calls Daybreak, and that the team used those models in work that ended with a vendor shipping fixes. The post presents that history as a reason the new agreement makes sense. A reason in a blog post is not a second announcement. This page will not describe the earlier flaws, the product they were in, or the method. The October 23 fact is the research agreement. The July fact is background the company chose to include, and it stops there.what a training cutoff does not mean.

Reuters has a second story in the lede

Reuters reports the collaboration accurately at the level of the parties and the topics: evaluate systems, find vulnerabilities, strengthen safeguards. It also sets the news against what it calls rising incidents of unwanted agent behavior, and against a separate episode it says OpenAI described. A wire story is allowed to supply context. Context is not an annex to the memorandum. A reader who came from Reuters and did not open OpenAI's post can think the NASK agreement is a response to that episode. OpenAI's post does not say that. This page will not insert the causal link Reuters leaves implicit.an evaluation a second lab can run.

Standards language is a policy sentence

OpenAI uses the post to repeat a call for countries to work on technical standards for frontier systems: how to evaluate capabilities, assess risk, test safeguards, and report incidents. NASK is offered as an example of a national institute whose daily work, the post says, is closer to how AI is actually used. A call for standards is not a standard. No draft number is in the post. The partnership is the news. The policy paragraph is the company's frame for why a Polish institute matters to a company in San Francisco.Phonon's file is a weight, this is not.

What NASK is, on these pages

OpenAI describes NASK as Poland's national research institute and describes the center as the national AI safety institute. NASK's announcement is in Polish and is the institute speaking for itself about the same partnership. Reuters calls NASK a national research institute. None of the three says NASK has acquired a veto over OpenAI's releases. A research agreement can be real and still leave every product decision where it was. This page does not invent governance the posts do not claim.what on-device inference keeps.

Who is in scope

NASK's researchers and OpenAI's safety staff, as the people the agreement names. Polish public-service users, only in the sense that NASK's existing work protects that environment, which is OpenAI's description and not a new service. Customers of a particular OpenAI product are not automatically in a changed terms-of-service. Readers who needed to know whether a model shipped today are in scope for the answer no. Asked directly, Is there a new model in this announcement? No. The posts describe a research collaboration. They do not name a ship date, a weight file, or a price. A safety institute agreement is not a product launch.Open Responses joined a foundation.

What stays on the record

October 23, 2026 OpenAI post and NASK announcement. Agreement to collaborate on safety, security, and technical research. NASK's center described as Poland's AI safety institute. Work described as evaluation, finding weaknesses, and safeguards. Quotes from Nielek and Duff Gordon, via OpenAI. July CERT Polska access mentioned as background, without the earlier flaws described here. No model release. Reuters as the wire account, including context this page does not treat as a term of the deal.

What would move the row

A joint evaluation with a published method, or a model card that names NASK as a tester, would be a later document. A text of the agreement with a term and a budget would be more than the two announcements. Until then the record is the two posts and Reuters's account of the signing. The short answer this page will stand on is this: OpenAI says it signed an agreement with NASK to collaborate on AI safety research. NASK's center would help evaluate systems and safeguards. The posts release no model. Reuters reports the signing and adds other safety disputes. Those disputes are Reuters's frame, not a term of the agreement.the advisory is the story.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. OpenAI
  2. NASK
  3. Reuters

Questions

Is there a new model in this announcement?

No. The posts describe a research collaboration. They do not name a ship date, a weight file, or a price. A safety institute agreement is not a product launch.

Who at NASK is named?

OpenAI quotes Radoslaw Nielek, NASK's director, on safeguards, synthetic content, and the need for institutes to work with developers. It quotes Tom Duff Gordon of OpenAI on using Polish researchers to test the company's assumptions. NASK's post is the institute's own account of the same signing.

What is the July background, without the flaws?

OpenAI says the agreement builds on earlier work with CERT Polska, which sits inside NASK, including access in July to cyber-capable models, and that a vendor later shipped fixes for issues that work helped find. This page does not describe those issues. Prior access is background in OpenAI's post. It is not the October 23 document.

What did Reuters add that the posts do not?

Reuters's account says the deal lands while OpenAI's safety practices are under scrutiny, and it refers to other incidents involving agents. Those sentences are Reuters's context. They are not clauses of the NASK agreement. This page will not treat them as findings of the partnership.

Is this the same story as a cloud reseller deal?

No. A separate brief filed this week is about selling OpenAI models through another company's catalog. This one is a research agreement with a Polish institute. Same company, different document.