OpenAI Signs AI Safety Research Agreement With Poland's NASK
The two will study how to evaluate AI systems, uncover their weak points, and strengthen safeguards. NASK's security team had already used OpenAI models to find six vulnerabilities in MikroTik RouterOS.
OpenAI announced on October 23 that it has signed an agreement with NASK, Poland’s national research institute, to explore opportunities for cooperation on AI safety, security, and technology research. On AI safety research specifically, the two organizations plan to work together directly, examining how to evaluate increasingly capable AI systems, find their weak points, and reinforce safeguards.
What they will study together
OpenAI’s counterpart is NASK’s AI safety research center, which serves as Poland’s AI Safety Institute (AISI). The scope of research the two sides will explore covers methods for evaluating AI systems, identifying hidden vulnerabilities, and strengthening safeguards. In other words, they intend to study together how to anticipate and address problems that may emerge as AI advances.
The announcement also highlighted an earlier instance of cooperation between the two in cybersecurity. According to OpenAI, CERT Polska, the security team housed within NASK, gained access to OpenAI’s cyber-capable models in July 2026 through Daybreak. CERT Polska put those models to work finding vulnerabilities in real-world software.
How they found six vulnerabilities
Using OpenAI models, CERT Polska discovered six vulnerabilities in MikroTik RouterOS. OpenAI explained that chaining some of those vulnerabilities together made it possible to take over the affected device without going through any authentication step. What stands out is less the number of weaknesses than the fact that, exploited in combination, they could hand over control of the device.
The vulnerabilities they found led to a vendor fix and follow-up verification. According to OpenAI, the manufacturer shipped fixes, and CERT Polska confirmed that the attacks it had observed were blocked. This is a case in which models helped identify the problem, and the team then verified that the fixes actually worked. For anyone running these devices, the key question is whether those fixes have been applied.
The significance of this agreement lies in the fact that NASK, with hands-on experience finding real vulnerabilities, is now joining forces with OpenAI on AI safety research. The earlier collaboration produced a concrete result: six vulnerabilities found, and attacks blocked once fixes were in place. What to watch now is which joint research projects follow, and what they deliver in terms of AI evaluation methods and improved safeguards.