OpenAI and Poland’s NASK partner to advance AI safety and security

OpenAI and NASK, Poland's National Research Institute, signed an agreement to explore joint work on AI safety, security and technical research. NASK's AI Safety Research Center acts as Poland's AI Safety Institute, and the two plan to work on evaluating more capable systems, finding vulnerabilities and strengthening safeguards.

It matters because frontier safety testing has mostly run through a handful of big-country institutes. Bringing in a mid-sized European state, with its own languages and its own threat picture, is how evaluation stops being a Silicon Valley-and-London conversation.

Let's be clear about what this is: an agreement to explore collaboration, not a published research program. There are no deliverables, timelines or funding figures in the announcement. Memorandums like this are easy to sign and easy to let go quiet, so we're holding our enthusiasm until we see work product.

That said, there's a real track record underneath it. CERT Polska, which sits inside NASK, got access to OpenAI's cyber-capable models in July through the Daybreak program. Per OpenAI, the team used them to help find six vulnerabilities in MikroTik RouterOS, including flaws that could be chained to take over devices without authentication. The vendor shipped fixes, and CERT Polska confirmed those fixes stopped the attacks it had observed. That's useful AI doing defensive work, with a public outcome, and we want more of it.

NASK also brings something OpenAI says it needs: expertise in multilingual risks, synthetic content detection and how AI gets misused against a national information ecosystem. Safety evaluations built mostly in English miss a lot. A Polish institute testing how models behave in Polish, against Polish public services, is a genuine contribution to the global picture.

The downside to name is independence. When the institutes meant to check frontier labs are also their research partners, the relationship can drift from oversight to collaboration. OpenAI frames these deals as feeding international technical standards, which we support. But a patchwork of bilateral agreements between one company and many governments is not the same as common rules, and it gives the lab a lot of say in what gets tested.

I'm watching for the first joint evaluation or published finding. That's what turns a signing into safety work.

OpenAI already works with the US Center for Advancing Standards and Innovation for Super Intelligence and the UK AI Security Institute, among others, and says those relationships have helped find and fix security vulnerabilities in AI agents. The company has called for the US to work with other countries on shared technical standards for frontier AI.