OpenAI partners with Poland's NASK on AI safety research

A new research partnership

OpenAI has entered into a formal agreement with Poland's NASK to conduct joint research on AI safety, security, and technical developments. This collaboration provides Polish public researchers with direct access to OpenAI's frontier models and their internal testing workflows. By sharing these tools, the companies aim to evaluate increasingly capable systems more effectively.

The partnership specifically involves NASK's AI Safety Research Center, which functions as the Polish Artificial Intelligence Safety Institute (AISI). This institute will work alongside OpenAI to identify potential risks and strengthen the safeguards surrounding new technologies. The goal is to ensure that as AI becomes more powerful, the measures to control it remain robust and effective.

OpenAI noted that many of the scientists who have led its own internal research efforts are Polish. This existing connection makes the partnership with NASK a logical progression for both parties. It allows Poland to contribute its research expertise to the broader international effort of making AI safer for everyone.

Testing and vulnerability discovery

The partnership is built on a practical testing loop that includes vulnerability discovery, vendor remediation, and validation. A real-world example of this occurred in July before the official agreement was signed. During that time, the incident response unit at NASK, known as CERT Polska, utilized OpenAI's cyber-focused models through a specific pathway called Daybreak.

Analysts used these models to identify six vulnerabilities in MikroTik RouterOS. These specific flaws were significant because they could be chained together to allow an attacker to take control of affected devices without needing any authentication. Once the flaws were discovered, MikroTik released the necessary fixes for the RouterOS software.

CERT Polska subsequently confirmed that the patches successfully prevented the recorded attacks against hardware that had not been updated. OpenAI intends to replicate this specific testing cycle with other national institutes. This process ensures that security flaws are found and fixed by vendors before they can be exploited by malicious actors.

Protecting public services

NASK is tasked with applying its safety research to protect Poland's digital information environment and public services. This includes a specific focus on detecting synthetic AI-generated content. Researchers will work to identify fake content and track how it might be used maliciously across different languages.

The collaboration aims to address emerging risks that could affect the general public. By understanding the multilingual aspects of synthetic content, researchers can better defend against disinformation. The partnership seeks to create a clearer picture of how malicious actors might use AI to manipulate information or disrupt services.

OpenAI emphasizes that drawing on expertise from outside its own organization is vital for challenging its own assumptions. By involving NASK, OpenAI can test its systems against a wider variety of perspectives. This collaborative approach is designed to strengthen the safeguards that protect users from the unintended consequences of advanced AI.

International safety standards

This agreement is part of a larger effort by OpenAI to coordinate with various international bodies. The company already works with the US Center for Advancing Standards and Innovation for Super Intelligence (CAISSI) and the UK AI Security Institute (AISI). These existing relationships have already helped identify and fix security vulnerabilities in AI agents.

OpenAI is currently advocating for international coordination regarding technical standards for frontier AI. The company believes that these standards should be developed in cooperation with other countries. This includes leveraging the expertise of various safety institutes and established standards bodies to create a unified approach to AI security.

The partnership with NASK opens a door for Poland to participate in these international efforts. By bringing Polish research into the fold, OpenAI hopes to create a more diverse and comprehensive safety framework. The ultimate goal is to establish a global standard for how frontier models are evaluated and deployed safely.

Shared research activities

OpenAI has outlined four specific shared activities that it intends to pursue with national institutes like NASK. These activities include capability evaluation, risk assessment, safeguard testing, and incident reporting. These pillars provide a structured way for researchers to analyze the potential impacts of new AI models.

Capability evaluation involves measuring what a model can actually do, while risk assessment looks at the potential harm those capabilities might cause. Safeguard testing ensures that the built-in protections are working as intended. Incident reporting provides a mechanism for sharing information when something goes wrong.

By focusing on these four areas, NASK and OpenAI can create a systematic way to monitor AI development. This structured approach helps ensure that safety is not an afterthought but is integrated into the research process. It allows for a proactive rather than reactive stance on AI security and public safety.