OpenAI and Poland's NASK Sign an AI Safety Agreement: What It Covers

OpenAI said on October 23, 2026 that it has signed an agreement with NASK, Poland's National Research Institute, to explore joint work on AI safety, security and technical research. The counterpart inside NASK is its AI Safety Research Center, which serves as the Polish AI Safety Institute. The official announcement is short on specifics, but it is a useful marker of how frontier labs are building ties with national safety bodies outside the US and UK.

This post covers what the announcement says, what it leaves out, and the concrete security work from CERT Polska that sits behind it. It also places the deal in a bigger fight: who gets to test frontier models before release, and under whose rules.

TL;DR: the OpenAI and NASK agreement at a glance

What the announcement actually says

OpenAI frames the deal as "an agreement to explore collaboration." The work will bring together expertise from both organizations "to better understand emerging risks, strengthen safeguards, and explore how advanced AI can contribute to a more secure and resilient digital ecosystem."

NASK's director, Radosław Nielek, is quoted describing the work as a natural step for an institute already active in AI safety research. He lists three strands: developing safeguards against emerging risks, detecting synthetic AI-generated content, and understanding both malicious uses and "the multilingual aspects of these challenges." That last point is the one most specific to Poland. Safety evaluations built mostly in English can miss failures that show up only in other languages, a gap that matters for any country trying to protect its own public services and information space.

Tom Duff Gordon, OpenAI's VP and Head of EMEA Policy, says Poland is "home to exceptional scientists and technical talent, many of whom have helped lead OpenAI's own research," and that outside expertise is needed "to challenge our assumptions, test our systems and strengthen safeguards."

Two facing shields, a gray and a green one, standing for OpenAI NASK partnership between a lab and a public safety institute

The announcement also ties the deal to a policy position. OpenAI says it has called for the US to work with other countries on international technical standards for frontier AI, drawing on safety institutes and standards bodies. Common ways of evaluating capabilities, assessing risks, testing safeguards and reporting incidents would let countries compare evidence and respond to risks that cross borders. In OpenAI's words, without national partners like NASK, "global AI safety standards may remain detached from the realities in which AI is actually used."

The concrete precedent: CERT Polska and six MikroTik flaws

The only measurable result cited is from cyber defense. In July, CERT Polska, which operates within NASK, gained access to OpenAI's cyber-capable models through Daybreak, the program OpenAI uses to give vetted defenders stronger security tooling. We covered how that program is tiered in our post on GPT-5.6-Cyber and the Daybreak red and blue tiers, and a customer-side example in Cloudflare's Daybreak vulnerability work.

OpenAI says the CERT Polska team used its models to help discover six vulnerabilities in MikroTik RouterOS, including flaws that could be combined to take control of affected devices without authentication. The vendor shipped fixes, and CERT Polska confirmed the fixes prevented the attacks it observed.

Independent reporting adds detail. Help Net Security reports that CERT Polska named the exploit chain MikroTrick. It describes three of the issues as a CVSS 9.2 SSH authentication bypass from incomplete RSA public-key validation, a CVSS 9.2 flaw in handling crafted usernames that can lift an SSH session to full administrator privileges, and a CVSS 8.8 bandwidth-test flaw that can leak kernel memory or crash a device. Exploitation was reportedly observed from at least September 2 against devices exposing SSH to public networks, and fixed RouterOS versions were released.

One caution on attribution: Help Net Security's write-up credits the discovery to GPT-5.5-cyber and GPT-5.6-sol models through a government collaboration program, while OpenAI's own post names Daybreak. We treat OpenAI's statement as the primary source for the program name and the outlet's report as secondary detail on the CVEs.

Why a Polish institute matters to a US lab

There are three practical reasons a lab like OpenAI wants this kind of agreement.

First, language and context coverage. Frontier-model safeguards are tested hardest in English. NASK's stated interest in multilingual misuse and synthetic-content detection fills a real gap, especially for disinformation aimed at a country on NATO's eastern flank. OpenAI's announcement mentions "protecting public services and Poland's digital information ecosystem" as the applied setting for NASK's research.

Second, defender access. The MikroTik case shows the model OpenAI prefers: give a trusted national CERT stronger tooling, let it find and report bugs, and use the results as evidence that gated access helps defense. That argument runs through OpenAI's Astra cybersecurity and Preparedness Framework disclosures, which describe why the strongest cyber capabilities are released to vetted groups first.

Third, politics. Government access to pre-release models has become contested. In September, US officials reportedly asked labs to hold new models back from the UK institute, an episode we followed in Anthropic's Mythos 5.1 and the UK AI Security Institute and in the White House reversal that restored UK access. A lab that signs agreements with several national institutes spreads its testing relationships and shows it is not tied to one government's gatekeeping.

What is not in the announcement

The gaps matter as much as the headline.

• No access terms. There is no statement on whether NASK gets pre-release models, which models, or under what confidentiality.
• No outputs promised. No joint papers, benchmarks, datasets or incident-sharing process are named.
• No money or timeline. Nothing says how long it runs or who pays.
• No independence language. The post does not say whether NASK can publish findings that OpenAI dislikes.
• No link to law. Poland is an EU member, so EU AI Act obligations apply there, but the announcement does not connect the agreement to the Act or to any Polish regulator's role.

These are normal for a first announcement of an "agreement to explore," and they are the questions to watch in the coming months. A useful test is whether a first output, such as a joint evaluation method or a shared incident report, appears within a year.

Red team probe pressing on a shell, representing vulnerability discovery by CERT Polska with OpenAI models

How this fits the wider push for AI security standards

The same week-to-week theme runs through several recent stories: governments want evidence they can trust, and labs want to show they cooperate. A bipartisan Senate letter to the White House on AI cyber standards, testing and auditing asked for common testing approaches in the US. OpenAI's own post points the same direction internationally, citing its September piece on building standards for the next phase of AI.

For builders, the through-line is that evaluation methods, incident reporting formats and safeguards testing are slowly becoming shared infrastructure, not private lab practice. Teams shipping agents should expect questions from customers and regulators about which external bodies have tested the underlying models.

What to do with this as a developer or security team

• Patch the underlying issue. If you run MikroTik gear, confirm you are on a fixed RouterOS release and that SSH is not exposed to the public internet. The advisories above list the fixed versions.
• Use gated defender programs if you qualify. Daybreak and similar programs exist so security teams can use stronger cyber tooling; see our Daybreak posts linked above for the tiers.
• Do not read the NASK deal as new product access. Nothing here changes ChatGPT, Codex or the API.
• Protect your own agents. If you deploy AI agents with real permissions, the same logic applies at your scale: test what they can do before they do it. AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.

Bottom line

The OpenAI and NASK agreement is a modest announcement with a strong supporting example. The example, six router vulnerabilities found with AI help and fixed before wider damage, is verified by both OpenAI and independent security reporting. The agreement itself is an intention to collaborate, with no public terms yet. Watch for what NASK publishes, which models it can test, and whether other EU institutes sign similar deals.

Related reading

• GPT-5.6-Cyber: OpenAI splits Daybreak into red and blue tiers
• Cloudflare and OpenAI Daybreak: vulnerability discovery and remediation
• OpenAI Astra, cybersecurity and the Preparedness Framework
• White House restores UK AI Security Institute access to frontier models
• Anthropic bars UK AI Security Institute from Mythos 5.1 testing
• Senate letter on AI cyber standards, testing and auditing

Details are accurate as of October 23, 2026. The agreement's terms may be clarified by OpenAI or NASK after publication.