OpenAI has added Poland to the governments whose safety researchers it works with directly. On October 23 the company said it had signed an agreement with NASK, the Polish National Research Institute, to explore cooperation on AI safety, security and technical research. Reuters reported the deal the same day.
The partner on the Polish side is NASK's AI Safety Research Center, which acts as Poland's official AI Safety Institute. According to OpenAI, the two will study how to evaluate increasingly capable models, look for vulnerabilities in them and work out stronger safeguards. NASK director Radosław Nielek described the center's existing research as covering protections against new risks, the detection of AI-generated content, malicious use of the technology and the problems that arise when models operate across many languages. He argued that safety institutes and frontier developers can only make real progress by working together.
For OpenAI, the agreement extends a pattern rather than starting one. The company says it already cooperates with CAISSI, the US government center that works on standards for superintelligence, the UK AI Security Institute and other public research bodies, and credits that work with helping it find and fix security flaws in AI agents. Tom Duff Gordon, who leads policy for Europe, the Middle East and Africa at OpenAI, pointed to Polish scientists who have held leading research roles at the company and said outside experts are needed to question its assumptions and test its systems.
There is also a track record on the cyber side. CERT Polska, the national computer emergency team that sits inside NASK, was given access in July to OpenAI's cyber-capable models through the Daybreak program. OpenAI says the team used them to uncover six vulnerabilities in MikroTik RouterOS, some of which could be chained to take over devices without logging in. MikroTik shipped patches, and CERT Polska confirmed they stopped the attacks it had been tracking.
OpenAI framed the deal as part of its push for shared international technical standards for frontier AI. It wants the US to work with other countries on common methods for measuring capabilities, assessing risk, testing safeguards and reporting incidents, so that evidence can be compared across borders. Reuters set the announcement against a run of disclosures by OpenAI, Anthropic and others about AI agents behaving in unwanted ways, which has pushed safety practices higher up the agenda.
What the agreement does not yet say matters as much as what it does. Both sides describe it as a framework to explore collaboration. Neither has published its duration, any funding, whether NASK will test OpenAI models before release, what level of model access it will get, or whether results will be made public. The claims about earlier results, including the RouterOS findings and the impact of OpenAI's work with US and UK institutes, come from OpenAI's own account.