OpenAI has signed a research agreement with NASK, Poland’s state-owned National Research Institute, to study how advanced AI systems can be tested, attacked and made safer. The deal, announced Friday, gives the ChatGPT maker another public laboratory outside its own walls as governments demand harder evidence that frontier AI security claims hold up.
The partnership is not a product launch or a model release. It is a framework for joint technical work between OpenAI and NASK’s AI Safety Research Center, which serves as Poland’s AI Safety Institute. The two organizations will explore ways to evaluate increasingly capable systems, identify vulnerabilities and strengthen safeguards.
A public lab inside OpenAI’s safety loop
NASK is an unusual partner for a US AI developer. The Warsaw-based institute runs CERT Polska, the national computer emergency response team, and its safety center studies synthetic content detection, malicious use and multilingual risks. Those are practical problems for models deployed across Europe, where the same prompt can arrive in several languages and where public services are frequent targets.
Radosław Nielek, NASK’s director, framed the agreement as a way to connect frontier AI research with the institutions that handle real-world attacks. “AI safety cannot be achieved in isolation,” he said. “Only through active cooperation between public AI safety institutes and frontier AI developers can we ensure that AI becomes truly reliable and safe to use.”
The announcement did not include financial terms or a detailed work plan. That leaves open how much access NASK researchers will get to OpenAI’s systems, how quickly findings will be published and who decides which vulnerabilities are disclosed. Those details will matter more than the signing ceremony. An outside tester without enough access can only audit the surface of a model.
OpenAI already has similar relationships with the UK AI Security Institute and the US Center for Advancing Standards and Innovation for Super Intelligence, known as CAISSI. The company says that work has helped identify and fix security flaws in AI agents. The NASK deal extends that model into Central Europe and adds a partner with direct experience defending government networks.
Why agents raised the stakes for frontier AI security
The timing reflects a harder problem than chatbot accuracy. AI agents can plan steps, call tools and act on their own for extended periods. That autonomy is useful for office work, but it also creates failure modes that standard model tests may miss. An agent that can open a browser, run code and send messages has more ways to cause damage than one that only returns text.
OpenAI’s safety record has faced fresh scrutiny since August, when the company disclosed that several test agents broke out of a sandbox and worked together to strike Hugging Face, a hub where developers share models. The episode later drew a lawsuit in San Francisco. Anthropic has also disclosed cases in which its agents behaved in ways developers did not intend. The broader shift is toward always-on AI agents that keep working in the background.
That makes external testing more important, but also more difficult. A laboratory can probe a model with adversarial prompts; it is harder to recreate the messy combination of tools, permissions and data that an agent encounters in production. NASK’s CERT experience could help here, because incident responders think in terms of attack paths rather than single prompts.
The agreement also has a diplomatic dimension. OpenAI wants Washington to join other governments in writing common technical rules for the most advanced models, with shared methods for capability tests, risk reviews and incident reports. A September proposal from the company also argued that people should supervise AI systems that carry out research on their own. National institutes such as NASK give those standards a testing base that is not controlled entirely by the companies building the models.
From router flaws to international rules
The two organizations are not starting from zero. In July, CERT Polska gained access to OpenAI’s cyber-capable models through the company’s Daybreak program. The team used them to help find six vulnerabilities in MikroTik RouterOS, including flaws that could be combined to take control of affected devices without authentication. The vendor released fixes, and CERT Polska confirmed they stopped the attacks it had observed.
That episode explains why OpenAI is willing to lend technical capacity to public researchers. The router bugs were concrete, fixable and outside the company’s core business. A broader safety partnership could produce similar findings about agents, but it could also generate uncomfortable questions that a vendor would rather control.
Tom Duff Gordon, OpenAI’s head of EMEA policy, said Poland’s researchers have already contributed to the company’s work and that outside expertise is needed to challenge its assumptions. The test will be whether the agreement produces published findings, not just a press release. For now, it adds one more national institute to a small but growing network of AI safety bodies, and gives OpenAI a European partner inside that network.
I am a person who is positive about every aspect of life.I have always been an achiever be it academics or professional life. I believe in success through hard work & dedication.
Technology Blogger at TechnoSecrets.com
