An international AI-safety agreement can sound abstract. The most tangible part of OpenAI's newly announced arrangement with Poland is what had already happened: a Polish cyber-response team used advanced models to discover six vulnerabilities in MikroTik RouterOS, and the vendor issued fixes. That experience offers a useful starting point for judging a broader research partnership announced on 23 October.[1]
OpenAI and Poland's NASK National Research Institute have signed an agreement to explore joint work on evaluating capable AI systems, identifying security weaknesses and strengthening safeguards. NASK's AI Safety Research Center serves as Poland's AI safety institute. The agreement describes areas for collaboration, not a purchase contract for a new AI system, a completed certification regime or a commitment to particular research findings.[1]
A patched router makes the promise testable
In July, CERT Polska, which operates within NASK, received access to OpenAI's cyber-capable models through the company's Daybreak initiative. The team identified six flaws in RouterOS, a widely deployed network operating system. According to OpenAI, some weaknesses could be combined to compromise affected devices without authentication. MikroTik released corrections and CERT Polska confirmed that the tested attacks were prevented after the fixes.[1]
This is a substantially stronger demonstration of utility than a generic claim that AI helps security. A real vendor product was examined, potential weaknesses were identified, and a corrective release followed. Yet six discovered vulnerabilities are not the same as six exploited breaches, and a corrected set of test cases is not proof that the product is free of all flaws. The relevant success is a bounded defensive workflow with a verifiable vendor response.[1]
Polish expertise matters for reasons beyond software patching. NASK works on public-sector security, synthetic-content detection and threats to the country's digital information environment. Evaluations in different languages and institutional settings can reveal failure patterns that a model's developer might not encounter in an English-language laboratory. A realistic test may ask whether an AI can detect a forged message circulating in Polish, or whether its recommendations remain sound under the constraints faced by local responders.[1]
Joint tests need independence, not just access
The same provider supplying a model may have a commercial incentive to emphasise good results. Public research institutions need the ability to define their own questions, record unsuccessful as well as successful tests and publish conclusions without the developer rewriting them. The new NASK agreement does not disclose an operational schedule or all contractual protections around publication rights, so readers should not assume those questions have been settled.[1]
There is an important distinction between evaluating a model's capability and assessing whether it is safe to deploy in a specific public service. A cybersecurity benchmark can show whether a model finds flaws. A procurement and operational assessment asks whether its access can be restricted, whether its suggestions are auditable, how false positives are handled and who owns a mistake. The two evaluations inform one another; neither replaces the other.[1]
The partnership also sits within a growing network of AI-safety institutions, including the UK AI Security Institute and public research programmes in the United States. Common tests could help compare systems, but scores will be misleading unless the participants agree on model versions, access to tools, threat assumptions, evaluator discretion and disclosure of incidents. An identical prompt does not create an identical test environment.[1]
What to watch next
The useful signs of progress will be public test protocols, independently reproducible findings, corrected real-world weaknesses and lessons that other security teams can apply. Joint research might also identify recurring failure modes in automated agents, such as unauthorised access, hidden communication channels or inability to stop a task safely. The announcement itself supplies no measured improvement in model behaviour since the agreement was signed.[1]
Poland's case illustrates a practical bargain at the centre of frontier-AI security. Developers can provide access to models that ordinary responders cannot easily evaluate. Independent institutions can bring local threat knowledge, defensive priorities and the discipline to test claims outside a vendor's own product demonstration. The initial six RouterOS fixes show what concrete collaboration can look like. The new agreement's value will be measured by the work that follows.[1]
Sources & further reading
- OpenAI and Poland’s NASK partner to advance AI safety and security OpenAI · 23 Oct 2026 · Accessed 26 Oct 2026